diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_1/index_tensor([8116, 8116])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_1/index_tensor([8116, 8116])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..184804ef817aa9a557fe4aa6b6125b0d12d13060 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_1/index_tensor([8116, 8116])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, top-down macro shot immerses the viewer in the complex textures of the damp, dark-gray asphalt surface, shimmering in the soft, diffuse light of an overcast sky. A small, shallow puddle forms within the slight depression, coated with a thin, shimmering film of oil, its surface swirling with a mesmerizing rainbow of vibrant hues of fuchsia, electric blue, and liquid gold. Suddenly, the film's delicate surface tension quietly breaks free; a rupture begins at the center and rapidly spreads outward, causing the psychedelic colors to instantly collapse and vanish. Simultaneously, a nearly invisible, fine mist erupts upward from the rupture point, suspended for a moment in the perfectly still, post-rain air before gracefully and completely dissipating, leaving only the dark, rain-slicked pavement behind, creating an atmosphere of fleeting beauty and tranquility.", + "seed": 21085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_1/index_8116_seed_21085_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.014162474265322089, + "vlm_logprob": -0.014162474265322089, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.014162474265322089, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04395033419132233 + ], + "opposite_logprob_tokens": [ + -4.29395055770874 + ], + "logprob_reward": -0.04395033419132233, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.04395033419132233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07955735176801682, + 0.039511993527412415, + 0.044269535690546036, + 0.04829278588294983, + 0.046079955995082855, + 0.10574423521757126, + 0.5134931206703186, + 0.07117348909378052, + 0.024332143366336823, + 0.027545390650629997 + ], + "credit_map": [ + 0.008118541911244392, + 0.0070519824512302876, + 0.010606965981423855, + 0.009751642122864723, + 0.005675254389643669, + 0.011855114251375198, + 0.011827372945845127, + 0.014670482836663723, + 0.0035714770201593637, + 0.004372645169496536, + 0.0026202534791082144, + 0.004717389587312937, + 0.004962013103067875, + 0.004307716153562069, + 0.009604910388588905, + 0.005355591420084238, + 0.00297657772898674, + 0.00739626307040453, + 0.005140678025782108, + 0.005892322864383459, + 0.009117967449128628, + 0.004325489979237318, + 0.005449322983622551, + 0.003970916848629713, + 0.002145062666386366, + 0.003100066212937236, + 0.003973045852035284, + 0.013362175785005093, + 0.005269552581012249, + 0.0027270005084574223, + 0.01387425884604454, + 0.0038416252937167883, + 0.0030270766001194715, + 0.0034013327676802874, + 0.007229159586131573, + 0.010102218948304653, + 0.0025165646802634, + 0.00463502760976553, + 0.007820025086402893, + 0.007348550017923117, + 0.006545898504555225, + 0.010803920216858387, + 0.010766573250293732, + 0.025732608512043953, + 0.014152135699987411, + 0.008290916681289673, + 0.019557559862732887, + 0.00989463645964861, + 0.019997652620077133, + 0.013416836969554424, + 0.018051480874419212, + 0.11186644434928894, + 0.016427533701062202, + 0.0119943143799901, + 0.22970347106456757, + 0.09203541278839111, + 0.005713697988539934, + 0.0066198864951729774, + 0.007620401680469513, + 0.017773354426026344, + 0.003826885251328349, + 0.009021598845720291, + 0.009500348940491676, + 0.011097319424152374, + 0.0018832803471013904, + 0.00448836013674736, + 0.0033856972586363554, + 0.0025169458240270615, + 0.002367883687838912, + 0.0026548877358436584, + 0.0037576400209218264, + 0.003277450567111373, + 0.002637171186506748, + 0.0029830574057996273, + 0.0034342787694185972, + 0.005177125334739685, + 0.002243744907900691, + 0.0035155864898115396, + 0.0037146196700632572, + 0.003839807352051139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.04395033419132233 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006775500252842903 + ], + "opposite_logprob_tokens": [ + -5.006775379180908 + ], + "logprob_reward": -0.006775500252842903, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006775500252842903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10618897527456284, + 0.03176484629511833, + 0.10896161943674088, + 0.1068287193775177, + 0.03065575286746025, + 0.019471516832709312, + 0.4331449866294861, + 0.10893809050321579, + 0.02854314260184765, + 0.025502236559987068 + ], + "credit_map": [ + 0.01206302922219038, + 0.006967648863792419, + 0.0286655705422163, + 0.01396435871720314, + 0.007345008663833141, + 0.013926634564995766, + 0.010405275970697403, + 0.01285146176815033, + 0.0030591182876378298, + 0.0035259760916233063, + 0.0029706424102187157, + 0.004900997970253229, + 0.00403676088899374, + 0.003271014429628849, + 0.0053826854564249516, + 0.004617656115442514, + 0.007232995238155127, + 0.013994657434523106, + 0.008617700077593327, + 0.013263956643640995, + 0.03378333896398544, + 0.013465278781950474, + 0.009682568721473217, + 0.00892113707959652, + 0.00448260921984911, + 0.007981917820870876, + 0.00950037781149149, + 0.03221029415726662, + 0.006460277363657951, + 0.005902037490159273, + 0.032226163893938065, + 0.008065052330493927, + 0.002548869699239731, + 0.003945885691791773, + 0.0044939289800822735, + 0.005391285754740238, + 0.002721847267821431, + 0.0031704746652394533, + 0.004950698930770159, + 0.0034327649045735598, + 0.0012868440244346857, + 0.002076831180602312, + 0.0024480081629008055, + 0.004964444786310196, + 0.0024464065209031105, + 0.0015982853947207332, + 0.002946709282696247, + 0.0017039892263710499, + 0.01685851626098156, + 0.012054286897182465, + 0.01594020426273346, + 0.09481972455978394, + 0.013219502754509449, + 0.010789957828819752, + 0.1921880543231964, + 0.07727479934692383, + 0.006737212650477886, + 0.006981364451348782, + 0.011635404080152512, + 0.02248983643949032, + 0.0039305961690843105, + 0.009306678548455238, + 0.03280681371688843, + 0.015050193294882774, + 0.002745716366916895, + 0.0023477186914533377, + 0.0042338743805885315, + 0.003412640420719981, + 0.0015598528552800417, + 0.004447540733963251, + 0.0052960580214858055, + 0.004499743692576885, + 0.002109315013512969, + 0.0029917703941464424, + 0.003654840402305126, + 0.004994179587811232, + 0.0015983552439138293, + 0.0027305088005959988, + 0.003064229153096676, + 0.004359040874987841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.006775500252842903 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000503770774230361 + ], + "opposite_logprob_tokens": [ + -7.6255035400390625 + ], + "logprob_reward": -0.000503770774230361, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.000503770774230361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15660490095615387, + 0.040062032639980316, + 0.029355254024267197, + 0.026748137548565865, + 0.03488457575440407, + 0.04020627588033676, + 0.5696017146110535, + 0.044677019119262695, + 0.04175795242190361, + 0.0161021389067173 + ], + "credit_map": [ + 0.01495146844536066, + 0.011957640759646893, + 0.016480453312397003, + 0.027545511722564697, + 0.013211362063884735, + 0.021569963544607162, + 0.016869787126779556, + 0.034018706530332565, + 0.0036203402560204268, + 0.005654813721776009, + 0.003082894952967763, + 0.005223568994551897, + 0.005542351398617029, + 0.004496823530644178, + 0.0075860582292079926, + 0.004855181556195021, + 0.0022469365503638983, + 0.004173442255705595, + 0.0032125303987413645, + 0.003495666431263089, + 0.007073304150253534, + 0.003529840847477317, + 0.003129206830635667, + 0.002494327025488019, + 0.001007750048302114, + 0.0013620018726214767, + 0.0014038186054676771, + 0.0073150163516402245, + 0.002180611714720726, + 0.0009362106211483479, + 0.010034289211034775, + 0.0025084351655095816, + 0.003045538207516074, + 0.003521238686516881, + 0.004671190865337849, + 0.0065651447512209415, + 0.001907562604174018, + 0.0029199684504419565, + 0.008287637494504452, + 0.003966293763369322, + 0.0026997514069080353, + 0.004001186229288578, + 0.004644796252250671, + 0.011504990980029106, + 0.0033695583697408438, + 0.002807177370414138, + 0.008082958869636059, + 0.0030958550050854683, + 0.022461360320448875, + 0.015175485983490944, + 0.020087487995624542, + 0.12604284286499023, + 0.016614433377981186, + 0.013158428482711315, + 0.25403425097465515, + 0.10202742367982864, + 0.003965593408793211, + 0.00332420552149415, + 0.005589056760072708, + 0.00659930519759655, + 0.0027280766516923904, + 0.005981489550322294, + 0.008885555900633335, + 0.007603737059980631, + 0.0038009078707545996, + 0.00388985313475132, + 0.005810929462313652, + 0.009124810807406902, + 0.002386299194768071, + 0.006336505524814129, + 0.004926861729472876, + 0.005481786094605923, + 0.0010914885206148028, + 0.0014118675608187914, + 0.0033058649860322475, + 0.002665986306965351, + 0.0008271995466202497, + 0.0012411260977387428, + 0.0028854976408183575, + 0.002673107897862792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.000503770774230361 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796163998544216 + ], + "opposite_logprob_tokens": [ + -4.634796142578125 + ], + "logprob_reward": -0.009796163998544216, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796163998544216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17884978652000427, + 0.10117358714342117, + 0.10266172885894775, + 0.06157999113202095, + 0.06484659016132355, + 0.03495710343122482, + 0.3528593182563782, + 0.059651654213666916, + 0.02490386925637722, + 0.018516361713409424 + ], + "credit_map": [ + 0.011125223711133003, + 0.018154483288526535, + 0.01996530219912529, + 0.04053938016295433, + 0.022709470242261887, + 0.018649311736226082, + 0.02579409070312977, + 0.021912546828389168, + 0.006233386695384979, + 0.013795492239296436, + 0.005709771066904068, + 0.01082424633204937, + 0.01795792393386364, + 0.006478590425103903, + 0.024999383836984634, + 0.015174810774624348, + 0.0073641324415802956, + 0.013410351239144802, + 0.00785653293132782, + 0.009598148986697197, + 0.032306279987096786, + 0.01197112537920475, + 0.011346080340445042, + 0.008809090591967106, + 0.0029470291920006275, + 0.0038290810771286488, + 0.003922546748071909, + 0.02002856694161892, + 0.0038756546564400196, + 0.003897004993632436, + 0.017820293083786964, + 0.0052598207257688046, + 0.004256365820765495, + 0.003598388284444809, + 0.006006558425724506, + 0.015253489837050438, + 0.004773205611854792, + 0.0041151512414216995, + 0.019702550023794174, + 0.0071408916264772415, + 0.0033252162393182516, + 0.0039779990911483765, + 0.006138316821306944, + 0.006040628533810377, + 0.0029612223152071238, + 0.003384161274880171, + 0.0054616802372038364, + 0.003667885437607765, + 0.014406424015760422, + 0.008690737187862396, + 0.013153799809515476, + 0.07497300952672958, + 0.010725491680204868, + 0.00869364570826292, + 0.15657564997673035, + 0.06564056873321533, + 0.003895910456776619, + 0.00532700028270483, + 0.005662776529788971, + 0.015844149515032768, + 0.0029483367688953876, + 0.006269786972552538, + 0.013642849400639534, + 0.0060608540661633015, + 0.002390366978943348, + 0.0021228413097560406, + 0.0038421410135924816, + 0.003934372216463089, + 0.001645896234549582, + 0.0036299959756433964, + 0.0046140579506754875, + 0.0027242011856287718, + 0.001102474401704967, + 0.003342116717249155, + 0.0023568172473460436, + 0.00381497316993773, + 0.0016497953329235315, + 0.001753325923345983, + 0.0021031966898590326, + 0.002393663628026843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.009796163998544216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786602109670639 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786602109670639, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786602109670639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17482690513134003, + 0.1490170955657959, + 0.08003948628902435, + 0.06780336797237396, + 0.08331701159477234, + 0.10016860067844391, + 0.07143403589725494, + 0.1296769231557846, + 0.11293837428092957, + 0.03077826090157032 + ], + "credit_map": [ + 0.010030192323029041, + 0.01540632639080286, + 0.012274166569113731, + 0.05442856624722481, + 0.020512504503130913, + 0.011080636642873287, + 0.021616829559206963, + 0.02947765216231346, + 0.011803855188190937, + 0.0193924643099308, + 0.011541332118213177, + 0.021212788298726082, + 0.02193286269903183, + 0.015920398756861687, + 0.029383791610598564, + 0.017829585820436478, + 0.00513227004557848, + 0.010437498800456524, + 0.005889935418963432, + 0.008030382916331291, + 0.025490164756774902, + 0.009227193892002106, + 0.009155861102044582, + 0.006676174234598875, + 0.0031062772031873465, + 0.005019993055611849, + 0.004211374558508396, + 0.02066361904144287, + 0.0064684669487178326, + 0.0030297101475298405, + 0.017308488488197327, + 0.007995427586138248, + 0.00555833475664258, + 0.005719327833503485, + 0.010008728131651878, + 0.018298709765076637, + 0.007283302955329418, + 0.007067319005727768, + 0.01815514639019966, + 0.011226131580770016, + 0.008902412839233875, + 0.011564595624804497, + 0.01687837392091751, + 0.02022230252623558, + 0.007410238031297922, + 0.00995204783976078, + 0.015042892657220364, + 0.010195726528763771, + 0.0027883523143827915, + 0.0040495507419109344, + 0.004783589392900467, + 0.0202320609241724, + 0.0029747101943939924, + 0.0033504341263324022, + 0.024932341650128365, + 0.008322994224727154, + 0.0073996237479150295, + 0.010319891385734081, + 0.010891550220549107, + 0.03655536100268364, + 0.0058065070770680904, + 0.00971534475684166, + 0.035254161804914474, + 0.013734460808336735, + 0.011339562945067883, + 0.009511767886579037, + 0.015937037765979767, + 0.020281238481402397, + 0.00648563914000988, + 0.01998562179505825, + 0.013852782547473907, + 0.015544714406132698, + 0.0027620606124401093, + 0.003916609566658735, + 0.004093796014785767, + 0.004491779953241348, + 0.0018901466391980648, + 0.005156372208148241, + 0.004015550948679447, + 0.0044519416987895966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.009786602109670639 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down macro shot immerses the viewer in the complex textures of the damp, dark-gray asphalt surface, shimmering in the soft, diffuse light of an overcast sky. A small, shallow puddle forms within the slight depression, coated with a thin, shimmering film of oil, its surface swirling with a mesmerizing rainbow of vibrant hues of fuchsia, electric blue, and liquid gold. Suddenly, the film's delicate surface tension quietly breaks free; a rupture begins at the center and rapidly spreads outward, causing the psychedelic colors to instantly collapse and vanish. Simultaneously, a nearly invisible, fine mist erupts upward from the rupture point, suspended for a moment in the perfectly still, post-rain air before gracefully and completely dissipating, leaving only the dark, rain-slicked pavement behind, creating an atmosphere of fleeting beauty and tranquility.", + "seed": 21086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_1/index_8116_seed_21086_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018876162869855762, + "vlm_logprob": -0.018876162869855762, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.018876162869855762, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056750934571027756 + ], + "opposite_logprob_tokens": [ + -3.806751012802124 + ], + "logprob_reward": -0.056750934571027756, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.056750934571027756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2087349146604538, + 0.05054106190800667, + 0.15213115513324738, + 0.15573927760124207, + 0.11816567182540894, + 0.0481610968708992, + 0.04160935804247856, + 0.05713748559355736, + 0.07743898779153824, + 0.09034102410078049 + ], + "credit_map": [ + 0.023223109543323517, + 0.026391634717583656, + 0.04834790527820587, + 0.020225098356604576, + 0.018700316548347473, + 0.03482910245656967, + 0.017019886523485184, + 0.01999785751104355, + 0.0053548808209598064, + 0.005071137100458145, + 0.007411273196339607, + 0.008226937614381313, + 0.006628409028053284, + 0.006630969699472189, + 0.006272337399423122, + 0.004945114720612764, + 0.01682770438492298, + 0.012899553403258324, + 0.04084977135062218, + 0.017522446811199188, + 0.01094949059188366, + 0.0194736048579216, + 0.01918981783092022, + 0.014418774284422398, + 0.01839551143348217, + 0.016134310513734818, + 0.03820251300930977, + 0.012414414435625076, + 0.012337137013673782, + 0.021125664934515953, + 0.022110560908913612, + 0.015019175596535206, + 0.012231986969709396, + 0.012637978419661522, + 0.014294777065515518, + 0.012540698051452637, + 0.0197307076305151, + 0.017307039350271225, + 0.01949777826666832, + 0.009924705140292645, + 0.005063361022621393, + 0.004760837648063898, + 0.00554434210062027, + 0.004002487752586603, + 0.012438464909791946, + 0.005002571269869804, + 0.007026362232863903, + 0.004322670865803957, + 0.004612214397639036, + 0.00430172448977828, + 0.009008022956550121, + 0.003976103384047747, + 0.0055737607181072235, + 0.0052261934615671635, + 0.004715905524790287, + 0.004195428919047117, + 0.005071835592389107, + 0.0074206190183758736, + 0.00581101281568408, + 0.004541402217000723, + 0.017478158697485924, + 0.0042830524034798145, + 0.008141616359353065, + 0.004389788955450058, + 0.008158200420439243, + 0.0072588506154716015, + 0.011647146195173264, + 0.00918627344071865, + 0.014957211911678314, + 0.007121137343347073, + 0.012959898449480534, + 0.0061502642929553986, + 0.011334871873259544, + 0.008859180845320225, + 0.014480143785476685, + 0.01129982154816389, + 0.010347988456487656, + 0.008308217860758305, + 0.018136484548449516, + 0.0075743175111711025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.056750934571027756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014236199669539928 + ], + "opposite_logprob_tokens": [ + -4.264235973358154 + ], + "logprob_reward": -0.014236199669539928, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014236199669539928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4949737787246704, + 0.11335604637861252, + 0.03320424631237984, + 0.03796659782528877, + 0.015354186296463013, + 0.035000160336494446, + 0.06214764714241028, + 0.08027014136314392, + 0.08153682202100754, + 0.04619026929140091 + ], + "credit_map": [ + 0.05531172454357147, + 0.04902665317058563, + 0.13593809306621552, + 0.04728621616959572, + 0.028544897213578224, + 0.08222393691539764, + 0.0428784042596817, + 0.05376390740275383, + 0.01219293661415577, + 0.008418317884206772, + 0.03126905858516693, + 0.011661646887660027, + 0.007185762282460928, + 0.01720512844622135, + 0.013015528209507465, + 0.012407679110765457, + 0.003560346085578203, + 0.0027510288637131453, + 0.005380232352763414, + 0.004224629141390324, + 0.0031414206605404615, + 0.0037401143927127123, + 0.007211500313133001, + 0.003194978227838874, + 0.0040418184362351894, + 0.003019117983058095, + 0.004007269628345966, + 0.0032001188956201077, + 0.005384083371609449, + 0.0034539534244686365, + 0.0115064000710845, + 0.003353837411850691, + 0.002251793397590518, + 0.0017922705737873912, + 0.00209984858520329, + 0.0012260464718565345, + 0.001892884960398078, + 0.0023886514827609062, + 0.0020861919037997723, + 0.0016164995031431317, + 0.004163109697401524, + 0.00417699571698904, + 0.006549712736159563, + 0.002155874390155077, + 0.005704679060727358, + 0.0053068590350449085, + 0.003603717777878046, + 0.0033392133191227913, + 0.007883652113378048, + 0.006206313613802195, + 0.006263932213187218, + 0.007618313189595938, + 0.00864684209227562, + 0.009713485836982727, + 0.010217064060270786, + 0.005598052870482206, + 0.007270841393619776, + 0.009405208751559258, + 0.009267361834645271, + 0.006279026158154011, + 0.023236451670527458, + 0.008074906654655933, + 0.01018553227186203, + 0.006550823338329792, + 0.00899538304656744, + 0.007758436258882284, + 0.008915601298213005, + 0.008433789014816284, + 0.017514878883957863, + 0.006849193945527077, + 0.01805276796221733, + 0.0050167785957455635, + 0.005107862874865532, + 0.004142977297306061, + 0.010860545560717583, + 0.004885678645223379, + 0.003930247854441404, + 0.005450726952403784, + 0.007509173359721899, + 0.004303058609366417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014236199669539928 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010519453790038824 + ], + "opposite_logprob_tokens": [ + -6.876051902770996 + ], + "logprob_reward": -0.0010519453790038824, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010519453790038824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17636027932167053, + 0.09362784773111343, + 0.043182622641325, + 0.07498106360435486, + 0.03808397799730301, + 0.03389371559023857, + 0.1470315009355545, + 0.12625084817409515, + 0.19155313074588776, + 0.0750349909067154 + ], + "credit_map": [ + 0.02524567022919655, + 0.01290294248610735, + 0.04040306806564331, + 0.02066986821591854, + 0.015641961246728897, + 0.022442888468503952, + 0.022578926756978035, + 0.01647494174540043, + 0.008870825171470642, + 0.006726495455950499, + 0.009385432116687298, + 0.014407293871045113, + 0.013321224600076675, + 0.008785772137343884, + 0.024436289444565773, + 0.007694512605667114, + 0.004848198499530554, + 0.003971046302467585, + 0.00832060445100069, + 0.004153503570705652, + 0.00460089510306716, + 0.00621863454580307, + 0.005725450813770294, + 0.005344287492334843, + 0.008337625302374363, + 0.006917546968907118, + 0.015548584051430225, + 0.007078766822814941, + 0.006486701779067516, + 0.009321415796875954, + 0.013896713964641094, + 0.007393708918243647, + 0.00413106894120574, + 0.003760307328775525, + 0.005394039675593376, + 0.0036767092533409595, + 0.0072016846388578415, + 0.0035663656890392303, + 0.007268991321325302, + 0.003084810683503747, + 0.004169442690908909, + 0.0032767883967608213, + 0.006093122065067291, + 0.0028661873657256365, + 0.005152201279997826, + 0.005234420765191317, + 0.003710720455273986, + 0.0033908335026353598, + 0.017964253202080727, + 0.01771557703614235, + 0.008530909195542336, + 0.013119403272867203, + 0.04163127765059471, + 0.017267126590013504, + 0.0186160858720541, + 0.012186877429485321, + 0.011798441410064697, + 0.015467892400920391, + 0.010680601932108402, + 0.011525139212608337, + 0.03624793514609337, + 0.012280643917620182, + 0.018274126574397087, + 0.00997606199234724, + 0.02052782103419304, + 0.01954801380634308, + 0.020277686417102814, + 0.022136857733130455, + 0.03972047567367554, + 0.018331272527575493, + 0.03892807289958, + 0.01208293717354536, + 0.00833083875477314, + 0.007264324463903904, + 0.014823850244283676, + 0.00984872318804264, + 0.008382217027246952, + 0.007980149239301682, + 0.01182094868272543, + 0.006583940703421831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0010519453790038824 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554306536912918 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.012554306536912918, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554306536912918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42492732405662537, + 0.1204601526260376, + 0.03888794407248497, + 0.0412595234811306, + 0.023523995652794838, + 0.03357411175966263, + 0.08744541555643082, + 0.16223347187042236, + 0.04447811469435692, + 0.02320992201566696 + ], + "credit_map": [ + 0.057408347725868225, + 0.041682809591293335, + 0.11769519001245499, + 0.031753916293382645, + 0.032951995730400085, + 0.05742275342345238, + 0.04569689556956291, + 0.0403154119849205, + 0.01500807423144579, + 0.009863514453172684, + 0.009236769750714302, + 0.021012037992477417, + 0.014144625514745712, + 0.014814975671470165, + 0.02574632503092289, + 0.010633816011250019, + 0.004850825760513544, + 0.003833893919363618, + 0.006411802489310503, + 0.005070073995739222, + 0.004233947955071926, + 0.004762782249599695, + 0.006250428501516581, + 0.0034741892013698816, + 0.004553706385195255, + 0.00406871410086751, + 0.006395020056515932, + 0.004104896914213896, + 0.00472557358443737, + 0.004750847816467285, + 0.009578853845596313, + 0.003081910777837038, + 0.0031208626460283995, + 0.0019700319971889257, + 0.0036061573773622513, + 0.002583869034424424, + 0.0029911217279732227, + 0.0031254172790795565, + 0.002972271526232362, + 0.003154264995828271, + 0.0034504386130720377, + 0.0031367989722639322, + 0.0043580010533332825, + 0.003063067328184843, + 0.00783674418926239, + 0.004442604724317789, + 0.00431601470336318, + 0.002970443805679679, + 0.01111180055886507, + 0.010282009840011597, + 0.005194409284740686, + 0.0069265360943973064, + 0.025872768834233284, + 0.009123601019382477, + 0.011752120219171047, + 0.0071821752935647964, + 0.013803064823150635, + 0.021011177450418472, + 0.011032797396183014, + 0.01383635401725769, + 0.05116528645157814, + 0.015226943418383598, + 0.022746065631508827, + 0.013411769643425941, + 0.0051126438193023205, + 0.004448135383427143, + 0.008885164745151997, + 0.003912433050572872, + 0.007117466069757938, + 0.005616546608507633, + 0.0054088677279651165, + 0.003976855427026749, + 0.0027034839149564505, + 0.002468528226017952, + 0.002612091600894928, + 0.0036431755870580673, + 0.003382287686690688, + 0.0034852123353630304, + 0.0027824020944535732, + 0.002132739871740341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.012554306536912918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787428192794323 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.009787428192794323, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787428192794323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18018488585948944, + 0.1436794400215149, + 0.13536299765110016, + 0.1624152660369873, + 0.08995513617992401, + 0.04905112832784653, + 0.04295075684785843, + 0.047674618661403656, + 0.09794501960277557, + 0.05078069865703583 + ], + "credit_map": [ + 0.01917862892150879, + 0.0170084647834301, + 0.021671373397111893, + 0.022435151040554047, + 0.01881943829357624, + 0.02037883549928665, + 0.043799784034490585, + 0.01689322292804718, + 0.016069242730736732, + 0.012197533622384071, + 0.01344789657741785, + 0.02031627483665943, + 0.01927834562957287, + 0.01243460550904274, + 0.036331064999103546, + 0.013604492880403996, + 0.016421442851424217, + 0.01377333328127861, + 0.01962214894592762, + 0.01583690568804741, + 0.012712890282273293, + 0.018573185428977013, + 0.024955011904239655, + 0.013468088582158089, + 0.021744374185800552, + 0.01872183568775654, + 0.034730955958366394, + 0.01309008989483118, + 0.013871538452804089, + 0.022903364151716232, + 0.022731643170118332, + 0.014621466398239136, + 0.009805099107325077, + 0.0100380415096879, + 0.00818245392292738, + 0.009013622999191284, + 0.016911009326577187, + 0.008597943000495434, + 0.02098538912832737, + 0.0064215874299407005, + 0.005765960086137056, + 0.0049024103209376335, + 0.005630320869386196, + 0.005123241804540157, + 0.009806495159864426, + 0.007783455774188042, + 0.005314784124493599, + 0.004724461585283279, + 0.005182785913348198, + 0.004266590811312199, + 0.005118833854794502, + 0.004744750447571278, + 0.005345022771507502, + 0.005854504648596048, + 0.008942863903939724, + 0.0034954058937728405, + 0.006460544653236866, + 0.005927402526140213, + 0.00514745619148016, + 0.00540332542732358, + 0.006977098528295755, + 0.008211031556129456, + 0.00496401684358716, + 0.004583746660500765, + 0.009696203283965588, + 0.009218486025929451, + 0.018529856577515602, + 0.008442932739853859, + 0.01935684308409691, + 0.012223117053508759, + 0.012590278871357441, + 0.007887309417128563, + 0.006092460826039314, + 0.004640480969101191, + 0.00976103637367487, + 0.00681165000423789, + 0.006276142317801714, + 0.007410848513245583, + 0.005195138044655323, + 0.00459294393658638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009787428192794323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down macro shot immerses the viewer in the complex textures of the damp, dark-gray asphalt surface, shimmering in the soft, diffuse light of an overcast sky. A small, shallow puddle forms within the slight depression, coated with a thin, shimmering film of oil, its surface swirling with a mesmerizing rainbow of vibrant hues of fuchsia, electric blue, and liquid gold. Suddenly, the film's delicate surface tension quietly breaks free; a rupture begins at the center and rapidly spreads outward, causing the psychedelic colors to instantly collapse and vanish. Simultaneously, a nearly invisible, fine mist erupts upward from the rupture point, suspended for a moment in the perfectly still, post-rain air before gracefully and completely dissipating, leaving only the dark, rain-slicked pavement behind, creating an atmosphere of fleeting beauty and tranquility.", + "seed": 21087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_1/index_8116_seed_21087_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.054519346263259647, + "vlm_logprob": -0.054519346263259647, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.054519346263259647, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22069504857063293 + ], + "opposite_logprob_tokens": [ + -1.7206950187683105 + ], + "logprob_reward": -0.22069504857063293, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.22069504857063293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07032462954521179, + 0.052230045199394226, + 0.027669155970215797, + 0.08874539285898209, + 0.2630675137042999, + 0.3108989894390106, + 0.02308029681444168, + 0.043308839201927185, + 0.08335116505622864, + 0.0373239628970623 + ], + "credit_map": [ + 0.0062840962782502174, + 0.0074379779398441315, + 0.0074110087007284164, + 0.012381047010421753, + 0.0064085484482347965, + 0.009557963348925114, + 0.01024281233549118, + 0.010601174086332321, + 0.008020971901714802, + 0.007072675973176956, + 0.005579961929470301, + 0.006326981354504824, + 0.0068033491261303425, + 0.006165692582726479, + 0.00713301170617342, + 0.005127401556819677, + 0.0036208259407430887, + 0.004782707430422306, + 0.002267630537971854, + 0.00355896377004683, + 0.0032234457321465015, + 0.003466187044978142, + 0.00334138423204422, + 0.003408009186387062, + 0.010066923685371876, + 0.018548620864748955, + 0.007257489953190088, + 0.0077127511613070965, + 0.021326936781406403, + 0.009155996143817902, + 0.0069840154610574245, + 0.007692658342421055, + 0.030902696773409843, + 0.03127113729715347, + 0.03398548439145088, + 0.0409567765891552, + 0.03731401264667511, + 0.025435838848352432, + 0.042463578283786774, + 0.02073797583580017, + 0.02357352338731289, + 0.08664969354867935, + 0.01833602786064148, + 0.021482476964592934, + 0.08686327189207077, + 0.029703451320528984, + 0.025771016255021095, + 0.018519548699259758, + 0.0023511210456490517, + 0.004021651577204466, + 0.0027922147419303656, + 0.002598194405436516, + 0.0030993898399174213, + 0.003099184948951006, + 0.0031182130333036184, + 0.00200032745487988, + 0.005789624527096748, + 0.007978111505508423, + 0.0035560866817831993, + 0.003651487175375223, + 0.007437921594828367, + 0.008249420672655106, + 0.0032837067265063524, + 0.003362478921189904, + 0.010216615162789822, + 0.012616191990673542, + 0.010101701132953167, + 0.007099386770278215, + 0.008673642762005329, + 0.011690075509250164, + 0.011735761538147926, + 0.011217794381082058, + 0.0033682638313621283, + 0.0071740178391337395, + 0.004349710885435343, + 0.004789481870830059, + 0.0031128900591284037, + 0.00640455074608326, + 0.0044705611653625965, + 0.003654486732557416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.22069504857063293 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026391128078103065 + ], + "opposite_logprob_tokens": [ + -3.65139102935791 + ], + "logprob_reward": -0.026391128078103065, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026391128078103065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33449235558509827, + 0.03820108249783516, + 0.041261497884988785, + 0.04416439309716225, + 0.10045542567968369, + 0.19115136563777924, + 0.07348258048295975, + 0.05648007616400719, + 0.0698852613568306, + 0.05042600259184837 + ], + "credit_map": [ + 0.03683801740407944, + 0.033443253487348557, + 0.0460711233317852, + 0.03842773661017418, + 0.028560485690832138, + 0.04310615733265877, + 0.04655968397855759, + 0.06148592010140419, + 0.004912139847874641, + 0.0041715907864272594, + 0.004490212071686983, + 0.005207845009863377, + 0.006467449478805065, + 0.004562324844300747, + 0.003979221917688847, + 0.00441029854118824, + 0.005666377488523722, + 0.005387567915022373, + 0.005383312236517668, + 0.005426968447864056, + 0.004976992961019278, + 0.005377102643251419, + 0.004543922375887632, + 0.00449925335124135, + 0.0048152687959373, + 0.006895090918987989, + 0.0043678502552211285, + 0.006608508061617613, + 0.006379816215485334, + 0.004896952770650387, + 0.006321310997009277, + 0.0038795978762209415, + 0.00810389406979084, + 0.017487183213233948, + 0.014124352484941483, + 0.01358332671225071, + 0.01170435268431902, + 0.013541441410779953, + 0.013489417731761932, + 0.008421453647315502, + 0.024988828226923943, + 0.039661504328250885, + 0.018491346389055252, + 0.014282834716141224, + 0.032859671860933304, + 0.023731235414743423, + 0.02019185945391655, + 0.016944067552685738, + 0.008459788747131824, + 0.010599741712212563, + 0.008256969973444939, + 0.007737886160612106, + 0.008508936502039433, + 0.010066086426377296, + 0.012835417874157429, + 0.007017751224339008, + 0.0035481948871165514, + 0.01698474772274494, + 0.004184926860034466, + 0.0024269847199320793, + 0.005116377957165241, + 0.018900539726018906, + 0.002839206950739026, + 0.0024790915194898844, + 0.009028210304677486, + 0.01084249746054411, + 0.0080202491953969, + 0.005652133841067553, + 0.0087104057893157, + 0.013575326651334763, + 0.006409365683794022, + 0.0076470691710710526, + 0.005464500747621059, + 0.009049107320606709, + 0.006496728863567114, + 0.006202524993568659, + 0.005386624019593, + 0.006654804572463036, + 0.006404857151210308, + 0.004766853991895914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.026391128078103065 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015209070406854153 + ], + "opposite_logprob_tokens": [ + -6.501521110534668 + ], + "logprob_reward": -0.0015209070406854153, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015209070406854153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14627477526664734, + 0.05063580349087715, + 0.03398206830024719, + 0.07215582579374313, + 0.1417185366153717, + 0.19494301080703735, + 0.04946436733007431, + 0.10331325232982635, + 0.11291854083538055, + 0.09459377080202103 + ], + "credit_map": [ + 0.016203945502638817, + 0.033269550651311874, + 0.014395834878087044, + 0.018357815220952034, + 0.015532171353697777, + 0.01793810911476612, + 0.012094588950276375, + 0.018482768908143044, + 0.004670377820730209, + 0.007908549159765244, + 0.004393203649669886, + 0.00702950544655323, + 0.008493109606206417, + 0.005575921852141619, + 0.006087520159780979, + 0.006477613467723131, + 0.004586467053741217, + 0.00461241789162159, + 0.0038820893969386816, + 0.004405670333653688, + 0.003926512785255909, + 0.00446800934150815, + 0.0038942648097872734, + 0.004206635523587465, + 0.007900726981461048, + 0.013725955970585346, + 0.005792363081127405, + 0.007835008203983307, + 0.015311640687286854, + 0.008206633850932121, + 0.006657411810010672, + 0.006726084742695093, + 0.01236227247864008, + 0.023055272176861763, + 0.022842885926365852, + 0.021594518795609474, + 0.015260572545230389, + 0.015917010605335236, + 0.021038638427853584, + 0.009647365659475327, + 0.014589478261768818, + 0.05223105475306511, + 0.011900577694177628, + 0.014119337312877178, + 0.05175754427909851, + 0.020113473758101463, + 0.01764385960996151, + 0.01258768793195486, + 0.005332615692168474, + 0.007185106631368399, + 0.0051586865447461605, + 0.0053587183356285095, + 0.005358646158128977, + 0.00621591554954648, + 0.0090052904561162, + 0.005849388428032398, + 0.007725242991000414, + 0.026867108419537544, + 0.008031750097870827, + 0.004471085965633392, + 0.010646969079971313, + 0.032854050397872925, + 0.005956436973065138, + 0.006760603282600641, + 0.014706123620271683, + 0.01536436565220356, + 0.012474928051233292, + 0.010431024245917797, + 0.011791073717176914, + 0.01927793025970459, + 0.015854304656386375, + 0.013018793426454067, + 0.01036115549504757, + 0.015336208045482635, + 0.013929000124335289, + 0.01214562263339758, + 0.009530365467071533, + 0.01267241220921278, + 0.011992564424872398, + 0.008626438677310944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0015209070406854153 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205524697899818 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.014205524697899818, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205524697899818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19199326634407043, + 0.05324776470661163, + 0.05104049667716026, + 0.07967972755432129, + 0.09348464757204056, + 0.24650134146213531, + 0.060911260545253754, + 0.057426415383815765, + 0.060951266437768936, + 0.10476376861333847 + ], + "credit_map": [ + 0.02359960414469242, + 0.022391337901353836, + 0.021596306934952736, + 0.025529325008392334, + 0.016327904537320137, + 0.022658606991171837, + 0.024997543543577194, + 0.03489263728260994, + 0.004203042946755886, + 0.007735009305179119, + 0.004154177848249674, + 0.009366180747747421, + 0.008194082416594028, + 0.006527407560497522, + 0.007690246216952801, + 0.005377620924264193, + 0.006603854708373547, + 0.0065431538969278336, + 0.005886984523385763, + 0.006621704436838627, + 0.005747923627495766, + 0.007664274889975786, + 0.005059418734163046, + 0.006913184653967619, + 0.006489685736596584, + 0.015560667030513287, + 0.006212869193404913, + 0.007051866035908461, + 0.021629072725772858, + 0.010074073448777199, + 0.005820480640977621, + 0.006841014605015516, + 0.008455983363091946, + 0.017188120633363724, + 0.014488921500742435, + 0.013016259297728539, + 0.008901931345462799, + 0.010282122530043125, + 0.014619782567024231, + 0.006531537510454655, + 0.017657887190580368, + 0.0684877559542656, + 0.01615224964916706, + 0.015226962976157665, + 0.06792408227920532, + 0.02554226666688919, + 0.021285666152834892, + 0.014224479906260967, + 0.007380762603133917, + 0.007651301100850105, + 0.009343757294118404, + 0.007722106296569109, + 0.006472337059676647, + 0.006370334420353174, + 0.010757739655673504, + 0.005212923977524042, + 0.007079611532390118, + 0.010534491389989853, + 0.005116371437907219, + 0.0037178448401391506, + 0.007947694510221481, + 0.01343464944511652, + 0.0041990140452980995, + 0.005396739579737186, + 0.008837424218654633, + 0.00866627600044012, + 0.008267647586762905, + 0.004392073955386877, + 0.007013981696218252, + 0.008662129752337933, + 0.007516753859817982, + 0.00759498355910182, + 0.012659894302487373, + 0.01631755940616131, + 0.014104287140071392, + 0.012844210490584373, + 0.011808999814093113, + 0.013083738274872303, + 0.012408881448209286, + 0.011536206118762493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.014205524697899818 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3257727026939392, + 0.05263656750321388, + 0.026618966832756996, + 0.06512732058763504, + 0.09184492379426956, + 0.17094974219799042, + 0.045841947197914124, + 0.04585742950439453, + 0.09778547286987305, + 0.07756491750478745 + ], + "credit_map": [ + 0.04257497936487198, + 0.026624489575624466, + 0.05399421602487564, + 0.03165901452302933, + 0.035853032022714615, + 0.03491326421499252, + 0.03470291197299957, + 0.06545082479715347, + 0.005601523444056511, + 0.009549492970108986, + 0.005287528969347477, + 0.006042607594281435, + 0.00955689512193203, + 0.006102320738136768, + 0.006066331174224615, + 0.004429865162819624, + 0.0035204521846026182, + 0.00335727003403008, + 0.002920606406405568, + 0.003754675853997469, + 0.0032217863481491804, + 0.003402457805350423, + 0.0027502444572746754, + 0.0036914728116244078, + 0.007426421623677015, + 0.011845668777823448, + 0.005845845676958561, + 0.006798483897000551, + 0.013837920501828194, + 0.007688021287322044, + 0.005605258978903294, + 0.006079693324863911, + 0.008398439735174179, + 0.012313351035118103, + 0.014616591855883598, + 0.017485132440924644, + 0.010051930323243141, + 0.008061680942773819, + 0.01448298804461956, + 0.006434808485209942, + 0.012673177756369114, + 0.04639735072851181, + 0.011658347211778164, + 0.013647806830704212, + 0.045778606086969376, + 0.014019631780683994, + 0.01707654632627964, + 0.009698264300823212, + 0.005617664195597172, + 0.005710866302251816, + 0.006473422981798649, + 0.0056343297474086285, + 0.005684737116098404, + 0.004500629846006632, + 0.006124739535152912, + 0.0060955556109547615, + 0.0037799652200192213, + 0.01071950327605009, + 0.0038098376244306564, + 0.00289327185600996, + 0.006031649652868509, + 0.012533631175756454, + 0.0029742531478405, + 0.003115316154435277, + 0.012654968537390232, + 0.014268835075199604, + 0.012859904207289219, + 0.009586429223418236, + 0.01096357125788927, + 0.017544208094477654, + 0.01066021341830492, + 0.009247342124581337, + 0.008955048397183418, + 0.010965865105390549, + 0.012024256400763988, + 0.008361715823411942, + 0.007520095910876989, + 0.012921147048473358, + 0.007456778548657894, + 0.009360008873045444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.009784122928977013 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, top-down macro shot immerses the viewer in the complex textures of the damp, dark-gray asphalt surface, shimmering in the soft, diffuse light of an overcast sky. A small, shallow puddle forms within the slight depression, coated with a thin, shimmering film of oil, its surface swirling with a mesmerizing rainbow of vibrant hues of fuchsia, electric blue, and liquid gold. Suddenly, the film's delicate surface tension quietly breaks free; a rupture begins at the center and rapidly spreads outward, causing the psychedelic colors to instantly collapse and vanish. Simultaneously, a nearly invisible, fine mist erupts upward from the rupture point, suspended for a moment in the perfectly still, post-rain air before gracefully and completely dissipating, leaving only the dark, rain-slicked pavement behind, creating an atmosphere of fleeting beauty and tranquility.", + "seed": 21088, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_1/index_8116_seed_21088_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015951806597877295, + "vlm_logprob": -0.015951806597877295, + "vlm_yes_no_margin": 5.274999952316284, + "vlm_token_credit": -0.015951806597877295, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05167977511882782 + ], + "opposite_logprob_tokens": [ + -4.051679611206055 + ], + "logprob_reward": -0.05167977511882782, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.05167977511882782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14026175439357758, + 0.20802883803844452, + 0.09789624810218811, + 0.10451386123895645, + 0.05526456981897354, + 0.08636678010225296, + 0.1278052181005478, + 0.07779400795698166, + 0.04375828430056572, + 0.05831042304635048 + ], + "credit_map": [ + 0.019767526537179947, + 0.018075132742524147, + 0.02700043097138405, + 0.013200019486248493, + 0.017529398202896118, + 0.011424871161580086, + 0.02040579915046692, + 0.012858587317168713, + 0.026390772312879562, + 0.0321328304708004, + 0.023411748930811882, + 0.030374472960829735, + 0.03459945321083069, + 0.018260547891259193, + 0.022640369832515717, + 0.020218662917613983, + 0.010112295858561993, + 0.012576454319059849, + 0.014721689745783806, + 0.012210327200591564, + 0.013033237308263779, + 0.011073127388954163, + 0.014022628776729107, + 0.010146496817469597, + 0.010261419229209423, + 0.014881741255521774, + 0.010332016274333, + 0.01966005750000477, + 0.013687670230865479, + 0.01030037458986044, + 0.013754772953689098, + 0.011635813862085342, + 0.006650570780038834, + 0.0063353790901601315, + 0.005651747342199087, + 0.008630293421447277, + 0.007231649477034807, + 0.006124814506620169, + 0.005967642180621624, + 0.00867247674614191, + 0.006708540488034487, + 0.01184719055891037, + 0.007163589354604483, + 0.013583830557763577, + 0.015421169809997082, + 0.009290768764913082, + 0.011887644417583942, + 0.01046405453234911, + 0.009814983233809471, + 0.019319657236337662, + 0.009916678071022034, + 0.028212932869791985, + 0.0227816179394722, + 0.009075811132788658, + 0.020089274272322655, + 0.008594270795583725, + 0.006367992144078016, + 0.012091409415006638, + 0.007631056010723114, + 0.013673418201506138, + 0.01187277864664793, + 0.006024674512445927, + 0.012207301333546638, + 0.007925386540591717, + 0.004149571526795626, + 0.0072316057048738, + 0.004256114829331636, + 0.0062099420465528965, + 0.008445528335869312, + 0.0038113242480903864, + 0.005744104739278555, + 0.003910096827894449, + 0.00590134272351861, + 0.008361389860510826, + 0.00793797243386507, + 0.01027251873165369, + 0.008629552088677883, + 0.0050696418620646, + 0.006524468306452036, + 0.005613546818494797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.05167977511882782 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002857889048755169 + ], + "opposite_logprob_tokens": [ + -5.877857685089111 + ], + "logprob_reward": -0.002857889048755169, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002857889048755169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4717612862586975, + 0.06762774288654327, + 0.1466827392578125, + 0.056086666882038116, + 0.03995858505368233, + 0.031122298911213875, + 0.0646556094288826, + 0.04309768229722977, + 0.021752355620265007, + 0.057255107909440994 + ], + "credit_map": [ + 0.09286509454250336, + 0.06831622123718262, + 0.11727435141801834, + 0.04310018569231033, + 0.036050815135240555, + 0.0351756326854229, + 0.03200974687933922, + 0.04696916788816452, + 0.01313814241439104, + 0.008658519014716148, + 0.011375874280929565, + 0.005963299423456192, + 0.007652789819985628, + 0.006327815819531679, + 0.006459973286837339, + 0.008051327429711819, + 0.01400501187890768, + 0.016017718240618706, + 0.016635987907648087, + 0.017034383490681648, + 0.017939237877726555, + 0.021412186324596405, + 0.01871451921761036, + 0.024923676624894142, + 0.006889261770993471, + 0.0066961608827114105, + 0.008001314476132393, + 0.008455729112029076, + 0.005791795905679464, + 0.00485646678134799, + 0.009253839030861855, + 0.006142086815088987, + 0.004490910097956657, + 0.004717148374766111, + 0.003711564466357231, + 0.005168354604393244, + 0.0051411516033113, + 0.006890490185469389, + 0.004136563744395971, + 0.0057023996487259865, + 0.0029566881712526083, + 0.004076395183801651, + 0.002764591481536627, + 0.004031579941511154, + 0.00446647172793746, + 0.003557346062734723, + 0.004328755661845207, + 0.004940467420965433, + 0.005026146303862333, + 0.008984943851828575, + 0.0074714371003210545, + 0.01321057602763176, + 0.0085241524502635, + 0.005879399832338095, + 0.009278850629925728, + 0.006280103698372841, + 0.004511184990406036, + 0.005869174376130104, + 0.0064698439091444016, + 0.008111327886581421, + 0.005255837459117174, + 0.0035788635723292828, + 0.005571624264121056, + 0.0037298232782632113, + 0.0025372393429279327, + 0.0030852125491946936, + 0.002687856089323759, + 0.004172338172793388, + 0.0023276268038898706, + 0.002030140021815896, + 0.0030575597193092108, + 0.0018543805927038193, + 0.005785307381302118, + 0.007427704520523548, + 0.00982334092259407, + 0.007687490899115801, + 0.007824154570698738, + 0.005801897030323744, + 0.006722935475409031, + 0.006182276178151369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.002857889048755169 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005022218101657927 + ], + "opposite_logprob_tokens": [ + -7.625502109527588 + ], + "logprob_reward": -0.0005022218101657927, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0005022218101657927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1414289027452469, + 0.21710693836212158, + 0.13203151524066925, + 0.10077841579914093, + 0.050318412482738495, + 0.056439485400915146, + 0.06483576446771622, + 0.07105820626020432, + 0.04224109649658203, + 0.12376128882169724 + ], + "credit_map": [ + 0.0257109347730875, + 0.02253618650138378, + 0.028675604611635208, + 0.01723525859415531, + 0.009992931969463825, + 0.009909192100167274, + 0.01614440605044365, + 0.011224386282265186, + 0.01791772060096264, + 0.027645278722047806, + 0.02776012010872364, + 0.035840995609760284, + 0.02882392518222332, + 0.018237249925732613, + 0.036966510117053986, + 0.023915136232972145, + 0.01194534171372652, + 0.012864530086517334, + 0.013067767024040222, + 0.012492038309574127, + 0.01554153673350811, + 0.022724734619259834, + 0.01515123713761568, + 0.028244316577911377, + 0.011326713487505913, + 0.011723734438419342, + 0.012895344756543636, + 0.014908931218087673, + 0.01153588481247425, + 0.011821016669273376, + 0.011541236191987991, + 0.015025552362203598, + 0.006304244045168161, + 0.006632158998399973, + 0.005083106923848391, + 0.008683302439749241, + 0.006154800299555063, + 0.00525428494438529, + 0.007625196129083633, + 0.004581314977258444, + 0.0045065018348395824, + 0.008045670576393604, + 0.004900188650935888, + 0.008568082936108112, + 0.009065099060535431, + 0.0058682807721197605, + 0.007685772608965635, + 0.007799890358000994, + 0.005260637030005455, + 0.009382843039929867, + 0.008478290401399136, + 0.012614186853170395, + 0.008643709123134613, + 0.005714587401598692, + 0.007906216196715832, + 0.0068352920934557915, + 0.006872072350233793, + 0.010693911463022232, + 0.007069026585668325, + 0.011878169141709805, + 0.012233602814376354, + 0.0068852901458740234, + 0.009084406308829784, + 0.006341724656522274, + 0.003704722970724106, + 0.006465760990977287, + 0.004017508123070002, + 0.007897771894931793, + 0.006404639687389135, + 0.0033468734472990036, + 0.006615572143346071, + 0.0037882423494011164, + 0.011550841853022575, + 0.018504289910197258, + 0.017469769343733788, + 0.017507387325167656, + 0.02343183383345604, + 0.010928765870630741, + 0.013587916269898415, + 0.01078047789633274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0005022218101657927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647376671433449 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647376671433449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647376671433449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2567695379257202, + 0.17365960776805878, + 0.12162536382675171, + 0.11938010901212692, + 0.04598388820886612, + 0.07907124608755112, + 0.09732604771852493, + 0.03847336769104004, + 0.03017016313970089, + 0.037540726363658905 + ], + "credit_map": [ + 0.03797626867890358, + 0.033913612365722656, + 0.042493969202041626, + 0.02333395555615425, + 0.028690211474895477, + 0.022136913612484932, + 0.04211461916565895, + 0.02610996924340725, + 0.019809769466519356, + 0.020829692482948303, + 0.022298073396086693, + 0.019466964527964592, + 0.03048649989068508, + 0.02007739432156086, + 0.020742710679769516, + 0.019948510453104973, + 0.012678953818976879, + 0.016523076221346855, + 0.011111349798738956, + 0.027822351083159447, + 0.015347317792475224, + 0.009353606961667538, + 0.019896483048796654, + 0.008892220444977283, + 0.011983362026512623, + 0.016160573810338974, + 0.009947793558239937, + 0.020114153623580933, + 0.014192353002727032, + 0.011605408973991871, + 0.020448366180062294, + 0.014928095042705536, + 0.004408024251461029, + 0.005167669150978327, + 0.005683014169335365, + 0.008234827779233456, + 0.00532526383176446, + 0.0051014721393585205, + 0.005626345984637737, + 0.006437270436435938, + 0.005573607515543699, + 0.010727119632065296, + 0.007410101126879454, + 0.013707930222153664, + 0.010671627707779408, + 0.007761084940284491, + 0.012223964557051659, + 0.01099581178277731, + 0.007137229200452566, + 0.014307447709143162, + 0.012005539610981941, + 0.018481750041246414, + 0.014311992563307285, + 0.008350087329745293, + 0.012253315187990665, + 0.010478693060576916, + 0.002806865144520998, + 0.005480597727000713, + 0.004178476985543966, + 0.006617615930736065, + 0.00465173926204443, + 0.003590828273445368, + 0.006815776228904724, + 0.004331468604505062, + 0.0030898703262209892, + 0.00451239850372076, + 0.0034640072844922543, + 0.006119830533862114, + 0.003929735627025366, + 0.0026738764718174934, + 0.004135573282837868, + 0.00224487017840147, + 0.0027534838300198317, + 0.005800073966383934, + 0.003929127939045429, + 0.008360130712389946, + 0.005399723537266254, + 0.0030429596081376076, + 0.0048782783560454845, + 0.0033769458532333374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008647376671433449 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a static top-down macro shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the surface dark-gray asphalt with damp textures visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small shallow puddle coated with a thin film of oil visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the oil film display swirling colors of fuchsia, electric blue, and liquid gold before breaking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the oil film rupture from the center and spread outward while a fine mist erupts upward and dissipates?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607177034020424 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.01607177034020424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607177034020424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14725592732429504, + 0.26897677779197693, + 0.11523491889238358, + 0.1276606321334839, + 0.09362319111824036, + 0.07152359187602997, + 0.06893686950206757, + 0.04185472056269646, + 0.036249104887247086, + 0.028684228658676147 + ], + "credit_map": [ + 0.02609071135520935, + 0.02516561932861805, + 0.031598713248968124, + 0.011341615580022335, + 0.014734772965312004, + 0.0126921022310853, + 0.011191902682185173, + 0.014440493658185005, + 0.03691335767507553, + 0.03196191415190697, + 0.0366668663918972, + 0.027987558394670486, + 0.04222266003489494, + 0.030448973178863525, + 0.034369274973869324, + 0.028406165540218353, + 0.013662002049386501, + 0.013071074150502682, + 0.017635276541113853, + 0.017199089750647545, + 0.013517703860998154, + 0.012133145704865456, + 0.014939483255147934, + 0.013077141717076302, + 0.012435374781489372, + 0.01692400872707367, + 0.011148852296173573, + 0.021143587306141853, + 0.013913895934820175, + 0.015097739174962044, + 0.018643464893102646, + 0.018353713676333427, + 0.00887440238147974, + 0.0095785241574049, + 0.010280870832502842, + 0.011828106828033924, + 0.009294381365180016, + 0.016723232343792915, + 0.009734324179589748, + 0.017309358343482018, + 0.005703276954591274, + 0.009084715507924557, + 0.006081799510866404, + 0.013115881010890007, + 0.011963809840381145, + 0.006833192892372608, + 0.0116190817207098, + 0.007121837232261896, + 0.005350073799490929, + 0.01046820916235447, + 0.007711478043347597, + 0.014319775626063347, + 0.011228783056139946, + 0.003772672498598695, + 0.010939694941043854, + 0.005146187264472246, + 0.003716531675308943, + 0.005781260784715414, + 0.0041822209022939205, + 0.010472484864294529, + 0.003318011062219739, + 0.0036101725418120623, + 0.006586370524019003, + 0.0041876682080328465, + 0.002831048797816038, + 0.005643456242978573, + 0.0027131487149745226, + 0.009770650416612625, + 0.004024970345199108, + 0.0020560729317367077, + 0.006865968927741051, + 0.002343790838494897, + 0.0026312165427953005, + 0.004313173238188028, + 0.003001742297783494, + 0.005718753673136234, + 0.0031993284355849028, + 0.002674506278708577, + 0.0033317639026790857, + 0.003813745453953743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.01607177034020424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_11/index_tensor([8020, 8020])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_11/index_tensor([8020, 8020])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fd620e53dc8b93d949020909e1696c231b3fd1ce --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_11/index_tensor([8020, 8020])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium shot captures an elderly Time Lord in a magnificent, soaring, domed council chamber. The image is captured in the soft, emotive style of Ghibli's hand-drawn animation, with delicate lines and saturated colors. The Time Lord is clad in a rich, draped velvet robe of crimson, trimmed with gold embroidery. A massive, rigid, fan-shaped collar of polished metal and dark fabric encircles his neck and rises behind his ears, framing his wrinkled, stern face. His brows are furrowed, his eyes narrowed in anger, and his lips set in a stern line, in the grip of a sharp accusation. He extends his ringed right hand, his index finger firmly and forcefully pointing at something off-screen. He leans forward slightly, his posture imbued with unquestionable authority. In the background, curved wooden benches are stacked one after another, huge stone pillars support the vaulted ceiling, and a beam of soft light pours down from the huge circular skylight in the center of the dome, illuminating the dust particles floating in the air, giving his gorgeous robe a soft glow, and casting a long, soft shadow behind him, creating a solemn and dramatic atmosphere of confrontation.", + "seed": 17270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_11/index_8020_seed_17270_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02375977858901024, + "vlm_logprob": -0.02375977858901024, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02375977858901024, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02101803757250309 + ], + "opposite_logprob_tokens": [ + -4.021018028259277 + ], + "logprob_reward": -0.02101803757250309, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02101803757250309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16179399192333221, + 0.04863908141851425, + 0.18732015788555145, + 0.03407793492078781, + 0.19866110384464264, + 0.05006183683872223, + 0.027578553184866905, + 0.02344454638659954, + 0.024675793945789337, + 0.24374698102474213 + ], + "credit_map": [ + 0.02940913662314415, + 0.019078083336353302, + 0.022713610902428627, + 0.018861418589949608, + 0.01585545390844345, + 0.020333409309387207, + 0.017213083803653717, + 0.0183298010379076, + 0.006197592709213495, + 0.0037480273749679327, + 0.01077304407954216, + 0.00563978124409914, + 0.0035147424787282944, + 0.004872066900134087, + 0.0045920140109956264, + 0.009301814250648022, + 0.07480636984109879, + 0.011061124503612518, + 0.02817908301949501, + 0.0124999163672328, + 0.007236560806632042, + 0.012219041585922241, + 0.011885476298630238, + 0.029432587325572968, + 0.004444492515176535, + 0.0040395306423306465, + 0.0034393956884741783, + 0.004892530385404825, + 0.003515903139486909, + 0.004695857875049114, + 0.00467783585190773, + 0.004372389055788517, + 0.025163894519209862, + 0.03871842473745346, + 0.020688721910119057, + 0.019468184560537338, + 0.034087173640728, + 0.020573247224092484, + 0.019656512886285782, + 0.020304935052990913, + 0.007911069318652153, + 0.00645996630191803, + 0.005632861517369747, + 0.00547600444406271, + 0.004453138913959265, + 0.00792614370584488, + 0.005465631373226643, + 0.006737019866704941, + 0.003485397668555379, + 0.006558071356266737, + 0.002146203303709626, + 0.003195263212546706, + 0.003083248622715473, + 0.00403787475079298, + 0.002304515102878213, + 0.002767981495708227, + 0.0033383388072252274, + 0.003409748896956444, + 0.002558181993663311, + 0.002937010722234845, + 0.002207699231803417, + 0.004045078065246344, + 0.003064412623643875, + 0.0018840753473341465, + 0.006769184023141861, + 0.001895539928227663, + 0.002664985368028283, + 0.0036264443770051003, + 0.001495671458542347, + 0.00386092159897089, + 0.0023627246264368296, + 0.0020003237295895815, + 0.05260155349969864, + 0.008469953201711178, + 0.09268215298652649, + 0.02842496894299984, + 0.005895203910768032, + 0.01471752766519785, + 0.028261534869670868, + 0.012694096192717552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.02101803757250309 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001953242812305689 + ], + "opposite_logprob_tokens": [ + -6.251953125 + ], + "logprob_reward": -0.001953242812305689, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001953242812305689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2261059433221817, + 0.07627412676811218, + 0.18999719619750977, + 0.07207535207271576, + 0.09021837264299393, + 0.05582789331674576, + 0.06358883529901505, + 0.04633297398686409, + 0.03699713945388794, + 0.14258219301700592 + ], + "credit_map": [ + 0.039349816739559174, + 0.031965263187885284, + 0.029266754165291786, + 0.022979481145739555, + 0.025025486946105957, + 0.029650714248418808, + 0.01963258534669876, + 0.02823583222925663, + 0.009913518093526363, + 0.009559550322592258, + 0.010769862681627274, + 0.011242418549954891, + 0.008623417466878891, + 0.010013068094849586, + 0.007825897075235844, + 0.008326392620801926, + 0.02467583492398262, + 0.009848804213106632, + 0.05591828376054764, + 0.013644542545080185, + 0.01053385529667139, + 0.01216723769903183, + 0.010337158106267452, + 0.05287148058414459, + 0.009591855108737946, + 0.008703057654201984, + 0.015032399445772171, + 0.007002958096563816, + 0.006973021198064089, + 0.009182034060359001, + 0.008411962538957596, + 0.007178067229688168, + 0.013694705441594124, + 0.015558047220110893, + 0.005901735741645098, + 0.010345707647502422, + 0.013152871280908585, + 0.009926073253154755, + 0.010792269371449947, + 0.010846961289644241, + 0.009446380659937859, + 0.010857276618480682, + 0.004360984079539776, + 0.004667052999138832, + 0.0077851214446127415, + 0.007836595177650452, + 0.004876972176134586, + 0.005997509695589542, + 0.010502331890165806, + 0.013759526424109936, + 0.005568388383835554, + 0.006730695255100727, + 0.010835434310138226, + 0.005974133033305407, + 0.005578739568591118, + 0.004639592487365007, + 0.013934251852333546, + 0.004515775013715029, + 0.003726125927641988, + 0.006421762052923441, + 0.0038528654258698225, + 0.007173598278313875, + 0.003031681524589658, + 0.0036769136786460876, + 0.004938962869346142, + 0.003512740135192871, + 0.004758770577609539, + 0.007512579206377268, + 0.0033124820329248905, + 0.005209015682339668, + 0.004569599404931068, + 0.0031829874496906996, + 0.02462274394929409, + 0.005239345598965883, + 0.058324750512838364, + 0.01604011096060276, + 0.0034637197386473417, + 0.012127717025578022, + 0.013410693034529686, + 0.00935311894863844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.001953242812305689 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000081241130829 + ], + "opposite_logprob_tokens": [ + -2.6950008869171143 + ], + "logprob_reward": -0.07000081241130829, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000081241130829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19656826555728912, + 0.051812466233968735, + 0.3164997100830078, + 0.13110393285751343, + 0.04746774956583977, + 0.030382219702005386, + 0.026038944721221924, + 0.07790885120630264, + 0.042268700897693634, + 0.07994920015335083 + ], + "credit_map": [ + 0.032061271369457245, + 0.022397154942154884, + 0.029996829107403755, + 0.01882462203502655, + 0.019170451909303665, + 0.02255440503358841, + 0.015513951890170574, + 0.03604957461357117, + 0.005850451532751322, + 0.004735658876597881, + 0.008691797032952309, + 0.006153798662126064, + 0.005201298277825117, + 0.006726082414388657, + 0.004699494689702988, + 0.009753885678946972, + 0.03426692262291908, + 0.01075702253729105, + 0.11141669005155563, + 0.0207156240940094, + 0.012494777329266071, + 0.016368422657251358, + 0.01617232896387577, + 0.09430792182683945, + 0.017458409070968628, + 0.015136994421482086, + 0.0271227415651083, + 0.014881013892591, + 0.015262712724506855, + 0.013499068096280098, + 0.012650211341679096, + 0.015092778950929642, + 0.006595924496650696, + 0.008956021629273891, + 0.004230631981045008, + 0.004895525984466076, + 0.0072457813657820225, + 0.005463320296257734, + 0.00507950596511364, + 0.005001034587621689, + 0.004925645422190428, + 0.005689335521310568, + 0.0027182414196431637, + 0.0030656822491437197, + 0.0030763279646635056, + 0.0035096234641969204, + 0.0035345074720680714, + 0.0038628571201115847, + 0.003339378861710429, + 0.005389214027673006, + 0.0022711551282554865, + 0.002582950284704566, + 0.0037276807706803083, + 0.004014759324491024, + 0.002250187797471881, + 0.002463617129251361, + 0.022088637575507164, + 0.01090950332581997, + 0.004497427027672529, + 0.007873223163187504, + 0.007672877050936222, + 0.013851718977093697, + 0.005344024859368801, + 0.005671441555023193, + 0.009987573139369488, + 0.0046132286079227924, + 0.00530760083347559, + 0.005131343845278025, + 0.003569834865629673, + 0.005619978066533804, + 0.004262953996658325, + 0.003776188939809799, + 0.013746768236160278, + 0.0029070405289530754, + 0.031127357855439186, + 0.009806135669350624, + 0.0022189028095453978, + 0.006212945561856031, + 0.008899500593543053, + 0.005030548200011253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.07000081241130829 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248814355581999 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005248814355581999, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248814355581999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12126182019710541, + 0.06004508584737778, + 0.40517187118530273, + 0.07529391348361969, + 0.03643367812037468, + 0.029686667025089264, + 0.03899925947189331, + 0.05510835722088814, + 0.04151324927806854, + 0.13648611307144165 + ], + "credit_map": [ + 0.020618587732315063, + 0.011736828833818436, + 0.01886070892214775, + 0.01236892119050026, + 0.01070399023592472, + 0.0141159538179636, + 0.011597362346947193, + 0.02125946618616581, + 0.007848456501960754, + 0.004572870209813118, + 0.011203798465430737, + 0.006365845445543528, + 0.003661497263237834, + 0.008403774350881577, + 0.006325648631900549, + 0.011663195677101612, + 0.14910057187080383, + 0.01267463993281126, + 0.14252012968063354, + 0.026152102276682854, + 0.00933592114597559, + 0.01692843995988369, + 0.023414138704538345, + 0.025045933201909065, + 0.013546559028327465, + 0.00706015108153224, + 0.00931963138282299, + 0.008404361084103584, + 0.005779176019132137, + 0.01244827639311552, + 0.007587725762277842, + 0.011148035526275635, + 0.0055900076404213905, + 0.003741851542145014, + 0.003785965731367469, + 0.004203027114272118, + 0.00489167170599103, + 0.004621254745870829, + 0.004353633150458336, + 0.005246266257017851, + 0.004040239378809929, + 0.004820271860808134, + 0.00307018356397748, + 0.003092102939262986, + 0.003148772055283189, + 0.004323659930378199, + 0.0030516679398715496, + 0.0041397688910365105, + 0.0054753003641963005, + 0.0067200022749602795, + 0.002910609357059002, + 0.003936683759093285, + 0.004801355302333832, + 0.008018859662115574, + 0.003317564260214567, + 0.0038188863545656204, + 0.015295637771487236, + 0.008074096404016018, + 0.003260642522946, + 0.004703771788626909, + 0.004992169328033924, + 0.009917272254824638, + 0.003690124023705721, + 0.005174639634788036, + 0.00651367474347353, + 0.0051564727909862995, + 0.0048309084959328175, + 0.005525728687644005, + 0.004530112724751234, + 0.0044456589967012405, + 0.0053080408833920956, + 0.005202650558203459, + 0.027648717164993286, + 0.004209179896861315, + 0.054478708654642105, + 0.015391968190670013, + 0.0029918125364929438, + 0.009581455960869789, + 0.01449527870863676, + 0.007689002901315689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.005248814355581999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577985793352127 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577985793352127, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577985793352127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3052239418029785, + 0.04272960498929024, + 0.3660927712917328, + 0.05454204976558685, + 0.05235786363482475, + 0.04888182505965233, + 0.01883222535252571, + 0.06249947473406792, + 0.021803533658385277, + 0.02703663893043995 + ], + "credit_map": [ + 0.0507269985973835, + 0.022301649674773216, + 0.08626052737236023, + 0.025973832234740257, + 0.023153645917773247, + 0.03266068920493126, + 0.024078229442238808, + 0.04006841778755188, + 0.006258747074753046, + 0.0028458740562200546, + 0.009425961412489414, + 0.0032801651395857334, + 0.0027788756415247917, + 0.006661189254373312, + 0.003915796522051096, + 0.007562997750937939, + 0.04372800141572952, + 0.008987256325781345, + 0.1509089469909668, + 0.023255739361047745, + 0.012061825953423977, + 0.01603570766746998, + 0.017441732808947563, + 0.09367355704307556, + 0.007263447158038616, + 0.007993211969733238, + 0.006470018997788429, + 0.005242241080850363, + 0.006499404553323984, + 0.0070186881348490715, + 0.004895346704870462, + 0.0091596944257617, + 0.007028026040643454, + 0.010591353289783001, + 0.004976990167051554, + 0.005611691623926163, + 0.009342136792838573, + 0.00458512594923377, + 0.0055979229509830475, + 0.004624622408300638, + 0.007550078444182873, + 0.00928010605275631, + 0.005587748251855373, + 0.004404295235872269, + 0.008140046149492264, + 0.005850846879184246, + 0.004148736596107483, + 0.003919969312846661, + 0.0028662681579589844, + 0.0041425894014537334, + 0.00199379725381732, + 0.0017729753162711859, + 0.002899473300203681, + 0.0019939816556870937, + 0.0014450475573539734, + 0.0017180939903482795, + 0.019035179167985916, + 0.007201173808425665, + 0.003485978115350008, + 0.006017740815877914, + 0.004639327526092529, + 0.013108713552355766, + 0.003870941698551178, + 0.0051404270343482494, + 0.003872502828016877, + 0.0030099109280854464, + 0.0023808469995856285, + 0.0025917470920830965, + 0.00255398522131145, + 0.002671030815690756, + 0.0022378219291567802, + 0.0024856894742697477, + 0.007344610523432493, + 0.0020791052374988794, + 0.006675354670733213, + 0.002499510534107685, + 0.0013854054268449545, + 0.0028641154058277607, + 0.0024387799203395844, + 0.0017497575609013438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.020577985793352127 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot captures an elderly Time Lord in a magnificent, soaring, domed council chamber. The image is captured in the soft, emotive style of Ghibli's hand-drawn animation, with delicate lines and saturated colors. The Time Lord is clad in a rich, draped velvet robe of crimson, trimmed with gold embroidery. A massive, rigid, fan-shaped collar of polished metal and dark fabric encircles his neck and rises behind his ears, framing his wrinkled, stern face. His brows are furrowed, his eyes narrowed in anger, and his lips set in a stern line, in the grip of a sharp accusation. He extends his ringed right hand, his index finger firmly and forcefully pointing at something off-screen. He leans forward slightly, his posture imbued with unquestionable authority. In the background, curved wooden benches are stacked one after another, huge stone pillars support the vaulted ceiling, and a beam of soft light pours down from the huge circular skylight in the center of the dome, illuminating the dust particles floating in the air, giving his gorgeous robe a soft glow, and casting a long, soft shadow behind him, creating a solemn and dramatic atmosphere of confrontation.", + "seed": 17271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_11/index_8020_seed_17271_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029439585004001855, + "vlm_logprob": -0.029439585004001855, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.029439585004001855, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01726420596241951 + ], + "opposite_logprob_tokens": [ + -4.142264366149902 + ], + "logprob_reward": -0.01726420596241951, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01726420596241951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.019981933757662773, + 0.019104884937405586, + 0.009209183976054192, + 0.007652735337615013, + 0.046349845826625824, + 0.2137833684682846, + 0.01916315034031868, + 0.04123866558074951, + 0.5013277530670166, + 0.12218845635652542 + ], + "credit_map": [ + 0.0023704874329268932, + 0.002203235635533929, + 0.003318528411909938, + 0.0020360478665679693, + 0.0025643869303166866, + 0.003129087621346116, + 0.0023624973837286234, + 0.0019976620096713305, + 0.002724296413362026, + 0.004287933465093374, + 0.0022006318904459476, + 0.0017080880934372544, + 0.0026071842294186354, + 0.002577661070972681, + 0.001352308550849557, + 0.001646780758164823, + 0.0010448753600940108, + 0.0012137520825490355, + 0.0011301763588562608, + 0.001249175169505179, + 0.0012195520102977753, + 0.001001694705337286, + 0.0013191381003707647, + 0.0010308196069672704, + 0.0013350901426747441, + 0.0009299471275880933, + 0.0012712107272818685, + 0.000528896227478981, + 0.0009456335683353245, + 0.0012030882062390447, + 0.0007644978468306363, + 0.0006743714911863208, + 0.015743078663945198, + 0.001615640940144658, + 0.002934743883088231, + 0.002442494733259082, + 0.0016319381538778543, + 0.01741194725036621, + 0.0028536091558635235, + 0.001716393162496388, + 0.08134512603282928, + 0.005512358155101538, + 0.007593971211463213, + 0.009685205295681953, + 0.006899663712829351, + 0.08291485905647278, + 0.010706108063459396, + 0.009126073680818081, + 0.006378850433975458, + 0.0008620963781140745, + 0.0007402602350339293, + 0.0010245537851005793, + 0.0012032483937218785, + 0.006924327928572893, + 0.0010323890019208193, + 0.0009974244749173522, + 0.015442756935954094, + 0.0011911033652722836, + 0.0013388149673119187, + 0.002193295629695058, + 0.0013075650203973055, + 0.01579086296260357, + 0.002041671657934785, + 0.0019325946923345327, + 0.18563447892665863, + 0.013775993138551712, + 0.01401625294238329, + 0.023143453523516655, + 0.01709846779704094, + 0.19812507927417755, + 0.024729067459702492, + 0.0248049795627594, + 0.044952645897865295, + 0.0039099776186048985, + 0.003431712742894888, + 0.005849594250321388, + 0.004659844096750021, + 0.04732794314622879, + 0.006181866861879826, + 0.0058748661540448666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.01726420596241951 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263755563646555 + ], + "opposite_logprob_tokens": [ + -5.255263805389404 + ], + "logprob_reward": -0.005263755563646555, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263755563646555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04540427029132843, + 0.050223108381032944, + 0.02552192658185959, + 0.010860028676688671, + 0.02126503735780716, + 0.25123536586761475, + 0.020319972187280655, + 0.08145549893379211, + 0.282664030790329, + 0.21105070412158966 + ], + "credit_map": [ + 0.008383936248719692, + 0.006927503738552332, + 0.006612352095544338, + 0.0026178359985351562, + 0.005033749155700207, + 0.008230616338551044, + 0.003636033507063985, + 0.003962249960750341, + 0.0069030895829200745, + 0.011970017105340958, + 0.0063617173582315445, + 0.003843015292659402, + 0.007460551802068949, + 0.006456100381910801, + 0.0035661966539919376, + 0.0036624225322157145, + 0.0038193247746676207, + 0.00392973143607378, + 0.004754501860588789, + 0.0016460069455206394, + 0.002689635381102562, + 0.005080003757029772, + 0.0017342159990221262, + 0.001868508872576058, + 0.0014859505463391542, + 0.002051673596724868, + 0.001638759975321591, + 0.0008807324920780957, + 0.0015011996729299426, + 0.0014908232260495424, + 0.0008836641791276634, + 0.0009272258030250669, + 0.004457125440239906, + 0.001409338554367423, + 0.0034183741081506014, + 0.0017898451769724488, + 0.0017887781141325831, + 0.005421741865575314, + 0.0014948963653296232, + 0.0014849401777610183, + 0.09538175165653229, + 0.006393512710928917, + 0.008875151164829731, + 0.011365138925611973, + 0.00806654803454876, + 0.09744185954332352, + 0.012747704982757568, + 0.010963737033307552, + 0.004763379227370024, + 0.0021853556390851736, + 0.0011641406454145908, + 0.0015378290554508567, + 0.0022701299749314785, + 0.005570448935031891, + 0.0014741354389116168, + 0.0013545551337301731, + 0.03034895472228527, + 0.0022770841605961323, + 0.0025487092789262533, + 0.0041929371654987335, + 0.0028878916054964066, + 0.0313945971429348, + 0.00404414813965559, + 0.003761191852390766, + 0.10468930006027222, + 0.007475828751921654, + 0.00802888534963131, + 0.013211973942816257, + 0.009849739260971546, + 0.11194079369306564, + 0.013891207985579967, + 0.013576344586908817, + 0.0766410082578659, + 0.007744473870843649, + 0.005709485150873661, + 0.010096268728375435, + 0.00878891721367836, + 0.08112052828073502, + 0.010780262760818005, + 0.010169782675802708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.005263755563646555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789543017745018 + ], + "opposite_logprob_tokens": [ + -2.5789542198181152 + ], + "logprob_reward": -0.0789543017745018, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789543017745018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028852878138422966, + 0.04475204274058342, + 0.02578648552298546, + 0.010996467433869839, + 0.06489734351634979, + 0.1722724437713623, + 0.018107209354639053, + 0.014878973364830017, + 0.5193668603897095, + 0.10008925199508667 + ], + "credit_map": [ + 0.004506914876401424, + 0.0034798055421561003, + 0.00430664699524641, + 0.0021670253481715918, + 0.0031476945150643587, + 0.005463141016662121, + 0.0029740938916802406, + 0.002807558048516512, + 0.00646046781912446, + 0.008549153804779053, + 0.007356966380029917, + 0.003443541470915079, + 0.005323320161551237, + 0.0056304349564015865, + 0.0029369480907917023, + 0.005051213316619396, + 0.004153077024966478, + 0.003186883870512247, + 0.004083284176886082, + 0.0028142989613115788, + 0.0018801994156092405, + 0.004951614420861006, + 0.0030981111340224743, + 0.0016190183814615011, + 0.0022787803318351507, + 0.0012165163643658161, + 0.0014045966090634465, + 0.0006835960666649044, + 0.0011312003480270505, + 0.0027056129183620214, + 0.0008627916104160249, + 0.0007133737090043724, + 0.022092243656516075, + 0.002831683261319995, + 0.0031250305473804474, + 0.003562904428690672, + 0.002714702393859625, + 0.024120457470417023, + 0.004014751873910427, + 0.0024355740752071142, + 0.0654347836971283, + 0.004598347935825586, + 0.006521528121083975, + 0.008220194838941097, + 0.0055936481803655624, + 0.06624798476696014, + 0.008326102048158646, + 0.007329858839511871, + 0.0028554925229400396, + 0.0026138832326978445, + 0.0017090715700760484, + 0.0019137702183797956, + 0.002262512920424342, + 0.0024648038670420647, + 0.00216963654384017, + 0.0021180403418838978, + 0.00231068953871727, + 0.002092806389555335, + 0.001370364218018949, + 0.0016559482319280505, + 0.002166090765967965, + 0.002014681464061141, + 0.0017793673323467374, + 0.0014890258898958564, + 0.19241194427013397, + 0.014363412745296955, + 0.014527590945363045, + 0.023768093436956406, + 0.01782049797475338, + 0.2048807144165039, + 0.02571076899766922, + 0.02588384598493576, + 0.037615928798913956, + 0.0030085269827395678, + 0.0030956892296671867, + 0.005078811198472977, + 0.003677312284708023, + 0.03884314000606537, + 0.0047791399993002415, + 0.003990705590695143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.0789543017745018 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624099031090736 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.007624099031090736, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624099031090736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04574435576796532, + 0.0313626267015934, + 0.053018808364868164, + 0.07283804565668106, + 0.11873404681682587, + 0.4222013056278229, + 0.018499048426747322, + 0.04452751204371452, + 0.13943064212799072, + 0.05364353582262993 + ], + "credit_map": [ + 0.009568107314407825, + 0.00570706557482481, + 0.007647277321666479, + 0.0033225943334400654, + 0.00494977505877614, + 0.005449498072266579, + 0.004423505626618862, + 0.004676534328609705, + 0.005210073199123144, + 0.003982692491263151, + 0.006161033641546965, + 0.0021885917522013187, + 0.003222227329388261, + 0.004327508620917797, + 0.002515417290851474, + 0.003755081444978714, + 0.009515299461781979, + 0.004960689693689346, + 0.010733889415860176, + 0.003431743010878563, + 0.004749922081828117, + 0.009351263754069805, + 0.004413513001054525, + 0.005862487945705652, + 0.01578453928232193, + 0.005183028988540173, + 0.015360564924776554, + 0.004042511340230703, + 0.004844161216169596, + 0.017343057319521904, + 0.005228404421359301, + 0.0050517781637609005, + 0.034227870404720306, + 0.00493065407499671, + 0.01976560801267624, + 0.0057428437285125256, + 0.004562779329717159, + 0.0348159596323967, + 0.008085645735263824, + 0.0066026863642036915, + 0.1595216542482376, + 0.01093338057398796, + 0.015230742283165455, + 0.018989576026797295, + 0.01394840981811285, + 0.1629219651222229, + 0.021471071988344193, + 0.019184529781341553, + 0.0056191496551036835, + 0.0013582094106823206, + 0.0010049876291304827, + 0.0011387437116354704, + 0.0011792882578447461, + 0.005867157131433487, + 0.0012709751026704907, + 0.0010605377610772848, + 0.015459960326552391, + 0.0019381289603188634, + 0.001639869064092636, + 0.0025472575798630714, + 0.002132516587153077, + 0.016198566183447838, + 0.0024840033147484064, + 0.0021272117737680674, + 0.0505291149020195, + 0.003769575385376811, + 0.004571324680000544, + 0.006872274447232485, + 0.005232442636042833, + 0.05440395325422287, + 0.0068696485832333565, + 0.007182301487773657, + 0.018384026363492012, + 0.002862591529265046, + 0.0016782116144895554, + 0.00269176228903234, + 0.0029183891601860523, + 0.01990830898284912, + 0.002739266026765108, + 0.002460979390889406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.007624099031090736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809156268835068 + ], + "opposite_logprob_tokens": [ + -3.2880916595458984 + ], + "logprob_reward": -0.03809156268835068, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809156268835068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034317292273044586, + 0.017022520303726196, + 0.09668398648500443, + 0.012773687019944191, + 0.04419422149658203, + 0.5341798067092896, + 0.007272804155945778, + 0.020843371748924255, + 0.21750059723854065, + 0.015211684629321098 + ], + "credit_map": [ + 0.004525887314230204, + 0.004477672278881073, + 0.006815964821726084, + 0.0023564079310745, + 0.0042482097633183, + 0.004535695072263479, + 0.003501596162095666, + 0.003855858463793993, + 0.0030120904557406902, + 0.002940161619335413, + 0.003176493337377906, + 0.0009441425208933651, + 0.002009038580581546, + 0.002083847764879465, + 0.0013921904610469937, + 0.0014645567862316966, + 0.010726754553616047, + 0.0064929393120110035, + 0.012685766443610191, + 0.01609557867050171, + 0.007561155594885349, + 0.012643945403397083, + 0.019934723153710365, + 0.010543129406869411, + 0.0021032821387052536, + 0.0015361852711066604, + 0.0023557625245302916, + 0.0009764657006599009, + 0.0017466628924012184, + 0.0018513378454372287, + 0.0012766282306984067, + 0.0009273625910282135, + 0.011367831379175186, + 0.0029987371526658535, + 0.007632396649569273, + 0.0026788273826241493, + 0.002462874399498105, + 0.011559400707483292, + 0.0032631652429699898, + 0.002230988582596183, + 0.20219393074512482, + 0.013614431954920292, + 0.019548214972019196, + 0.023987410590052605, + 0.01729602925479412, + 0.20622512698173523, + 0.027024343609809875, + 0.02429032139480114, + 0.0013809287920594215, + 0.001490525552071631, + 0.0005534837255254388, + 0.0004938928759656847, + 0.0008605571929365396, + 0.0013932237634435296, + 0.0006433372036553919, + 0.0004568551084958017, + 0.007443564012646675, + 0.0008215096895582974, + 0.0007173140766099095, + 0.0011068175081163645, + 0.0009783556452021003, + 0.007709377445280552, + 0.0010768674546852708, + 0.0009895650437101722, + 0.08062096685171127, + 0.005863001104444265, + 0.006112001836299896, + 0.0100787915289402, + 0.007548377383500338, + 0.08611996471881866, + 0.01070069894194603, + 0.010456792078912258, + 0.004893868695944548, + 0.0009051834349520504, + 0.0009452725062146783, + 0.0009476471459493041, + 0.0008600361179560423, + 0.004818625282496214, + 0.0008438311633653939, + 0.000997220166027546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.03809156268835068 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot captures an elderly Time Lord in a magnificent, soaring, domed council chamber. The image is captured in the soft, emotive style of Ghibli's hand-drawn animation, with delicate lines and saturated colors. The Time Lord is clad in a rich, draped velvet robe of crimson, trimmed with gold embroidery. A massive, rigid, fan-shaped collar of polished metal and dark fabric encircles his neck and rises behind his ears, framing his wrinkled, stern face. His brows are furrowed, his eyes narrowed in anger, and his lips set in a stern line, in the grip of a sharp accusation. He extends his ringed right hand, his index finger firmly and forcefully pointing at something off-screen. He leans forward slightly, his posture imbued with unquestionable authority. In the background, curved wooden benches are stacked one after another, huge stone pillars support the vaulted ceiling, and a beam of soft light pours down from the huge circular skylight in the center of the dome, illuminating the dust particles floating in the air, giving his gorgeous robe a soft glow, and casting a long, soft shadow behind him, creating a solemn and dramatic atmosphere of confrontation.", + "seed": 17272, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_11/index_8020_seed_17272_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05311605464667082, + "vlm_logprob": -0.05311605464667082, + "vlm_yes_no_margin": 3.299999976158142, + "vlm_token_credit": -0.05311605464667082, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04899502545595169 + ], + "opposite_logprob_tokens": [ + -3.173995018005371 + ], + "logprob_reward": -0.04899502545595169, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04899502545595169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12200925499200821, + 0.12251175194978714, + 0.10069254785776138, + 0.09791438281536102, + 0.1296827793121338, + 0.09576217085123062, + 0.1256064772605896, + 0.0417146198451519, + 0.08232720196247101, + 0.08177880197763443 + ], + "credit_map": [ + 0.009007962420582771, + 0.029460016638040543, + 0.009507780894637108, + 0.024115506559610367, + 0.015859223902225494, + 0.010388839989900589, + 0.01629769802093506, + 0.0073722247034311295, + 0.007661785930395126, + 0.007651233579963446, + 0.00408518873155117, + 0.036180030554533005, + 0.01260096300393343, + 0.008248623460531235, + 0.0386204831302166, + 0.007463441696017981, + 0.00916579831391573, + 0.01837272383272648, + 0.00840746145695448, + 0.015844179317355156, + 0.012374777346849442, + 0.010077926330268383, + 0.013641725294291973, + 0.012807952240109444, + 0.008091019466519356, + 0.0200079046189785, + 0.008101983927190304, + 0.016123002395033836, + 0.008717228658497334, + 0.011809330433607101, + 0.012566580437123775, + 0.012497331015765667, + 0.006232190877199173, + 0.013186747208237648, + 0.0037452552933245897, + 0.015863483771681786, + 0.03874240070581436, + 0.03512212261557579, + 0.010734082199633121, + 0.006056496873497963, + 0.011306962929666042, + 0.013512231409549713, + 0.009120701812207699, + 0.014806780032813549, + 0.014691448770463467, + 0.011644238606095314, + 0.012324785813689232, + 0.008355028927326202, + 0.010995479300618172, + 0.019212692975997925, + 0.010468238964676857, + 0.01466106902807951, + 0.021946139633655548, + 0.018490225076675415, + 0.01612226665019989, + 0.013710370287299156, + 0.00496536772698164, + 0.008215322159230709, + 0.0023928810842335224, + 0.004497850779443979, + 0.009073038585484028, + 0.004860735032707453, + 0.004429489374160767, + 0.003279934637248516, + 0.007106132339686155, + 0.010454810224473476, + 0.007280928548425436, + 0.010994448326528072, + 0.016703344881534576, + 0.011473022401332855, + 0.011329633183777332, + 0.00698487926274538, + 0.006735942792147398, + 0.014954179525375366, + 0.0069945515133440495, + 0.010241350159049034, + 0.014536322094500065, + 0.01240469329059124, + 0.00989675335586071, + 0.006015019956976175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.04899502545595169 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018215281888842583 + ], + "opposite_logprob_tokens": [ + -4.018215179443359 + ], + "logprob_reward": -0.018215281888842583, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018215281888842583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08838315308094025, + 0.30107489228248596, + 0.03691047802567482, + 0.0364922471344471, + 0.2147992104291916, + 0.03116513043642044, + 0.04272199794650078, + 0.09859737753868103, + 0.04448378458619118, + 0.10537175834178925 + ], + "credit_map": [ + 0.006460235919803381, + 0.010367553681135178, + 0.008049385622143745, + 0.012130005285143852, + 0.019959423691034317, + 0.012657703831791878, + 0.012745977379381657, + 0.0060128578916192055, + 0.01941210776567459, + 0.013984675519168377, + 0.007846253924071789, + 0.05112803354859352, + 0.07624419778585434, + 0.06079027056694031, + 0.05672628805041313, + 0.014943054877221584, + 0.0025453572161495686, + 0.005364151671528816, + 0.0033090284559875727, + 0.004858773667365313, + 0.00577869126573205, + 0.005308927036821842, + 0.004896146710962057, + 0.004849398974329233, + 0.0033393464982509613, + 0.0069021740928292274, + 0.002498905174434185, + 0.004834756720811129, + 0.003767163958400488, + 0.006886571645736694, + 0.004292515106499195, + 0.003970813937485218, + 0.00788681861013174, + 0.011128897778689861, + 0.0037601131480187178, + 0.008586946874856949, + 0.07926152646541595, + 0.08820591121912003, + 0.007815816439688206, + 0.008153161965310574, + 0.0040492950938642025, + 0.003042023628950119, + 0.0023812896106392145, + 0.0046370625495910645, + 0.003988438751548529, + 0.005723923444747925, + 0.0036367876455187798, + 0.003706305520609021, + 0.003450952237471938, + 0.008100945502519608, + 0.0033245489466935396, + 0.005909200292080641, + 0.007299486082047224, + 0.005115865729749203, + 0.006021158304065466, + 0.0034998345654457808, + 0.009502250701189041, + 0.018910424783825874, + 0.008568395860493183, + 0.01465856283903122, + 0.016053322702646255, + 0.009937545284628868, + 0.012851137667894363, + 0.008115731179714203, + 0.0046529690735042095, + 0.007475142367184162, + 0.0038077733479440212, + 0.006030089221894741, + 0.007626155856996775, + 0.006084505468606949, + 0.005281078163534403, + 0.0035260680597275496, + 0.008695492520928383, + 0.023424236103892326, + 0.006095510441809893, + 0.010538192465901375, + 0.022094478830695152, + 0.014440466649830341, + 0.01156411599367857, + 0.008519251830875874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.018215281888842583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277730882167816 + ], + "opposite_logprob_tokens": [ + -2.01777720451355 + ], + "logprob_reward": -0.14277730882167816, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14277730882167816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06845894455909729, + 0.1857968419790268, + 0.033708054572343826, + 0.025910336524248123, + 0.21014557778835297, + 0.0750516951084137, + 0.06421682983636856, + 0.08991002291440964, + 0.2256098836660385, + 0.0211917944252491 + ], + "credit_map": [ + 0.005344451405107975, + 0.008326169103384018, + 0.005588608793914318, + 0.010580507107079029, + 0.014885708689689636, + 0.009321453049778938, + 0.010302750393748283, + 0.004109307192265987, + 0.011601435020565987, + 0.010387680493295193, + 0.005137181840837002, + 0.03665047883987427, + 0.04012390971183777, + 0.029325872659683228, + 0.041147004812955856, + 0.011423289775848389, + 0.0027972469106316566, + 0.005724933929741383, + 0.0030063095036894083, + 0.004758537746965885, + 0.004530624952167273, + 0.004211819265037775, + 0.005295044742524624, + 0.0033835398498922586, + 0.0020697624422609806, + 0.004839854780584574, + 0.0022432839032262564, + 0.0030285578686743975, + 0.002416018396615982, + 0.004009857773780823, + 0.003363753901794553, + 0.003939249087125063, + 0.007678656838834286, + 0.012271241284906864, + 0.0039013989735394716, + 0.012251044623553753, + 0.07552710175514221, + 0.08125254511833191, + 0.009979235008358955, + 0.0072843641974031925, + 0.00797835923731327, + 0.006424459628760815, + 0.0032710430677980185, + 0.009807204827666283, + 0.010464916937053204, + 0.01882203109562397, + 0.007340868469327688, + 0.010942822322249413, + 0.006624503992497921, + 0.007277510594576597, + 0.005736113991588354, + 0.007196532096713781, + 0.010639671236276627, + 0.010014249943196774, + 0.007265621330589056, + 0.009462629444897175, + 0.00833280198276043, + 0.017991703003644943, + 0.007251907140016556, + 0.012308117933571339, + 0.018112273886799812, + 0.009286779910326004, + 0.01054180134087801, + 0.006084646563977003, + 0.019832996651530266, + 0.03132006153464317, + 0.020723972469568253, + 0.02941422536969185, + 0.04482026770710945, + 0.03247138112783432, + 0.027174539864063263, + 0.019852451980113983, + 0.0016679558902978897, + 0.0047728680074214935, + 0.0010817332658916712, + 0.0021297885105013847, + 0.004789065569639206, + 0.0029178608674556017, + 0.002179131843149662, + 0.0016533943125978112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.14277730882167816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537707574665546 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537707574665546, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537707574665546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19574078917503357, + 0.061156436800956726, + 0.03124941699206829, + 0.1574571579694748, + 0.11672163754701614, + 0.11609820276498795, + 0.07829601317644119, + 0.06233668327331543, + 0.14953245222568512, + 0.03141122683882713 + ], + "credit_map": [ + 0.020577322691679, + 0.018739238381385803, + 0.01604781672358513, + 0.02448931708931923, + 0.031493641436100006, + 0.03324345499277115, + 0.025734124705195427, + 0.025415880605578423, + 0.004256097599864006, + 0.008392165414988995, + 0.0035934364423155785, + 0.006519456394016743, + 0.014491411857306957, + 0.012686323374509811, + 0.006365326698869467, + 0.004852221813052893, + 0.0022882057819515467, + 0.00548844039440155, + 0.0030751763842999935, + 0.0051365564577281475, + 0.003926399629563093, + 0.0035710709635168314, + 0.004453766625374556, + 0.0033098002895712852, + 0.013141494244337082, + 0.019270725548267365, + 0.012278705835342407, + 0.02304091304540634, + 0.01594620756804943, + 0.020104235038161278, + 0.02143603004515171, + 0.03223884105682373, + 0.0046953437849879265, + 0.006706365384161472, + 0.0024535132106393576, + 0.005483563523739576, + 0.04124680534005165, + 0.04517588019371033, + 0.005376524291932583, + 0.005583651829510927, + 0.012801258824765682, + 0.013429482467472553, + 0.006950072944164276, + 0.01388130709528923, + 0.017052361741662025, + 0.026289209723472595, + 0.010791725479066372, + 0.014902791939675808, + 0.007630772422999144, + 0.010181646794080734, + 0.0066177197732031345, + 0.00859047006815672, + 0.013663961552083492, + 0.011963131837546825, + 0.009457690641283989, + 0.010190622881054878, + 0.005533824674785137, + 0.011730967089533806, + 0.003907998092472553, + 0.007052604574710131, + 0.015973534435033798, + 0.007032955531030893, + 0.006500724703073502, + 0.004604075103998184, + 0.013435207307338715, + 0.021022198721766472, + 0.013759253546595573, + 0.019716139882802963, + 0.02902001328766346, + 0.02170983888208866, + 0.017875730991363525, + 0.012994085438549519, + 0.0026463400572538376, + 0.00566442497074604, + 0.003114889608696103, + 0.003528535133227706, + 0.005126771982759237, + 0.005222422070801258, + 0.003703253110870719, + 0.002404590602964163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.012537707574665546 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305494949221611 + ], + "opposite_logprob_tokens": [ + -3.1680550575256348 + ], + "logprob_reward": -0.04305494949221611, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305494949221611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25317808985710144, + 0.18629074096679688, + 0.03810311481356621, + 0.06228543817996979, + 0.06308889389038086, + 0.09700930863618851, + 0.09118504822254181, + 0.05472750589251518, + 0.11468671262264252, + 0.03944512829184532 + ], + "credit_map": [ + 0.027522599324584007, + 0.02234385907649994, + 0.0206342451274395, + 0.03215501457452774, + 0.03183702751994133, + 0.05839800462126732, + 0.028340540826320648, + 0.0319468192756176, + 0.013300021179020405, + 0.009630699642002583, + 0.006253063678741455, + 0.011823371984064579, + 0.059260234236717224, + 0.056307900696992874, + 0.011437776498496532, + 0.018277689814567566, + 0.002786749042570591, + 0.00693663302809, + 0.004711045417934656, + 0.00500663835555315, + 0.004938654135912657, + 0.004649376962333918, + 0.0050852554850280285, + 0.003988764714449644, + 0.0052419682033360004, + 0.010522034950554371, + 0.0038616154342889786, + 0.006275375839322805, + 0.009722994640469551, + 0.013984305784106255, + 0.005818959791213274, + 0.006858182139694691, + 0.0024247628170996904, + 0.004386238753795624, + 0.0017805652460083365, + 0.004984158091247082, + 0.020950617268681526, + 0.022136038169264793, + 0.003906821832060814, + 0.0025196915958076715, + 0.011357223615050316, + 0.011145347729325294, + 0.0046460614539682865, + 0.01161088328808546, + 0.013463360257446766, + 0.023877128958702087, + 0.008351996541023254, + 0.012557316571474075, + 0.007636530324816704, + 0.012920868583023548, + 0.0070521957240998745, + 0.009188764728605747, + 0.019924288615584373, + 0.012905571609735489, + 0.01107127033174038, + 0.01048556249588728, + 0.004181781318038702, + 0.009922733530402184, + 0.003235589014366269, + 0.005196481477469206, + 0.014777792617678642, + 0.006773893255740404, + 0.005662561859935522, + 0.004976675380021334, + 0.01049571018666029, + 0.01595977693796158, + 0.010634036734700203, + 0.014979260973632336, + 0.022298386320471764, + 0.016590775921940804, + 0.013868516311049461, + 0.009860261343419552, + 0.0029829160775989294, + 0.008587407879531384, + 0.002881028689444065, + 0.004280859604477882, + 0.008236614987254143, + 0.0045800888910889626, + 0.004647743422538042, + 0.003248468739911914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.04305494949221611 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot captures an elderly Time Lord in a magnificent, soaring, domed council chamber. The image is captured in the soft, emotive style of Ghibli's hand-drawn animation, with delicate lines and saturated colors. The Time Lord is clad in a rich, draped velvet robe of crimson, trimmed with gold embroidery. A massive, rigid, fan-shaped collar of polished metal and dark fabric encircles his neck and rises behind his ears, framing his wrinkled, stern face. His brows are furrowed, his eyes narrowed in anger, and his lips set in a stern line, in the grip of a sharp accusation. He extends his ringed right hand, his index finger firmly and forcefully pointing at something off-screen. He leans forward slightly, his posture imbued with unquestionable authority. In the background, curved wooden benches are stacked one after another, huge stone pillars support the vaulted ceiling, and a beam of soft light pours down from the huge circular skylight in the center of the dome, illuminating the dust particles floating in the air, giving his gorgeous robe a soft glow, and casting a long, soft shadow behind him, creating a solemn and dramatic atmosphere of confrontation.", + "seed": 17273, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_11/index_8020_seed_17273_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.033632824290543796, + "vlm_logprob": -0.033632824290543796, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.033632824290543796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023420369252562523 + ], + "opposite_logprob_tokens": [ + -3.8984203338623047 + ], + "logprob_reward": -0.023420369252562523, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023420369252562523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20673102140426636, + 0.12853817641735077, + 0.07985767722129822, + 0.04530094936490059, + 0.04593747854232788, + 0.052181635051965714, + 0.0454128123819828, + 0.048757582902908325, + 0.2171618640422821, + 0.13012079894542694 + ], + "credit_map": [ + 0.029037974774837494, + 0.03148698806762695, + 0.01683966815471649, + 0.02737291529774666, + 0.026611724868416786, + 0.04370566084980965, + 0.020668035373091698, + 0.011008054949343204, + 0.028070740401744843, + 0.013930422253906727, + 0.013219594955444336, + 0.01477535255253315, + 0.012460982427001, + 0.02480144239962101, + 0.01229799259454012, + 0.008981652557849884, + 0.012919551692903042, + 0.007706018164753914, + 0.008392076008021832, + 0.007034360431134701, + 0.008878769353032112, + 0.017769820988178253, + 0.007912903092801571, + 0.00924417283385992, + 0.005867790896445513, + 0.006088598631322384, + 0.004427944775670767, + 0.004015340935438871, + 0.008301397785544395, + 0.009848300367593765, + 0.004342542961239815, + 0.0024090316146612167, + 0.005432118661701679, + 0.004694199655205011, + 0.007238300982862711, + 0.005348979029804468, + 0.005744237452745438, + 0.00655304454267025, + 0.003535853000357747, + 0.007390744984149933, + 0.007630486972630024, + 0.004051378928124905, + 0.011562692001461983, + 0.006192388013005257, + 0.004299372900277376, + 0.005806560628116131, + 0.00442295428365469, + 0.008215798996388912, + 0.010504408739507198, + 0.004221427254378796, + 0.006834167055785656, + 0.005045361816883087, + 0.0031102958600968122, + 0.008358711376786232, + 0.0034291925840079784, + 0.003909248858690262, + 0.008247614838182926, + 0.00641605956479907, + 0.005814190488308668, + 0.0054625533521175385, + 0.006511436775326729, + 0.006296374835073948, + 0.006446077022701502, + 0.003563277656212449, + 0.03094753623008728, + 0.01949343830347061, + 0.023830721154808998, + 0.03878900408744812, + 0.017338374629616737, + 0.04312065243721008, + 0.02597746253013611, + 0.017664674669504166, + 0.019670382142066956, + 0.021886935457587242, + 0.014217542484402657, + 0.014080353081226349, + 0.019093720242381096, + 0.01748691126704216, + 0.013925798237323761, + 0.009759158827364445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.023420369252562523 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963627714663744 + ], + "opposite_logprob_tokens": [ + -5.1309638023376465 + ], + "logprob_reward": -0.005963627714663744, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963627714663744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13365691900253296, + 0.15422043204307556, + 0.09195280820131302, + 0.061763063073158264, + 0.04695877805352211, + 0.060552433133125305, + 0.0540732704102993, + 0.07116558402776718, + 0.08042735606431961, + 0.24522936344146729 + ], + "credit_map": [ + 0.01809433102607727, + 0.019547801464796066, + 0.01244876254349947, + 0.015477454289793968, + 0.015333175659179688, + 0.03060377761721611, + 0.012254893779754639, + 0.00989671889692545, + 0.021008701995015144, + 0.02252192609012127, + 0.01911463961005211, + 0.02236347459256649, + 0.01762198470532894, + 0.02657138928771019, + 0.012178627774119377, + 0.012839690782129765, + 0.014333902858197689, + 0.009992017410695553, + 0.010431068949401379, + 0.006718005053699017, + 0.010136482305824757, + 0.025092249736189842, + 0.006398273166269064, + 0.008850803598761559, + 0.0061284746043384075, + 0.006975595839321613, + 0.008435635827481747, + 0.008334181271493435, + 0.012158775702118874, + 0.008436803705990314, + 0.006678170990198851, + 0.004615424666553736, + 0.006743363104760647, + 0.006411840207874775, + 0.004914437886327505, + 0.004005359020084143, + 0.007478958927094936, + 0.006215756293386221, + 0.004649085458368063, + 0.006539974361658096, + 0.008850909769535065, + 0.006204669829457998, + 0.007074151188135147, + 0.005330486688762903, + 0.006899955216795206, + 0.011598470620810986, + 0.005340071860700846, + 0.009253714233636856, + 0.007920158095657825, + 0.007111164275556803, + 0.0071677593514323235, + 0.008100359700620174, + 0.005742146633565426, + 0.006664951331913471, + 0.006610197015106678, + 0.00475652888417244, + 0.013170856982469559, + 0.008403537794947624, + 0.007078848779201508, + 0.007358546834439039, + 0.009767549112439156, + 0.009779638610780239, + 0.008765705861151218, + 0.006840893998742104, + 0.01729891076683998, + 0.007581653073430061, + 0.007194214500486851, + 0.005327608902007341, + 0.005908242426812649, + 0.02605469524860382, + 0.005130271427333355, + 0.005931761581450701, + 0.047133710235357285, + 0.03113255649805069, + 0.017732270061969757, + 0.02359667979180813, + 0.028020324185490608, + 0.06561125069856644, + 0.020301971584558487, + 0.011700601316988468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.005963627714663744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901481330394745 + ], + "opposite_logprob_tokens": [ + -2.464014768600464 + ], + "logprob_reward": -0.08901481330394745, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901481330394745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17810958623886108, + 0.25504443049430847, + 0.10959817469120026, + 0.0598582960665226, + 0.05187408626079559, + 0.04074181616306305, + 0.03692052140831947, + 0.05230993777513504, + 0.12194757163524628, + 0.09359563142061234 + ], + "credit_map": [ + 0.02103792317211628, + 0.023239046335220337, + 0.022045811638236046, + 0.024736233055591583, + 0.021377651020884514, + 0.031734734773635864, + 0.020009122788906097, + 0.013929038308560848, + 0.05186353251338005, + 0.0361824557185173, + 0.01859637349843979, + 0.02663755603134632, + 0.03765669837594032, + 0.032387785613536835, + 0.026115095242857933, + 0.02560490369796753, + 0.016557853668928146, + 0.011913314461708069, + 0.012864413671195507, + 0.008383066393435001, + 0.009876414202153683, + 0.03062768094241619, + 0.009323802776634693, + 0.010051622055470943, + 0.007769709452986717, + 0.005926762241870165, + 0.011142916046082973, + 0.0074522229842841625, + 0.0037252248730510473, + 0.012904955074191093, + 0.0057455021888017654, + 0.005190995987504721, + 0.00829022191464901, + 0.005837190430611372, + 0.007421679794788361, + 0.0033119458239525557, + 0.0066907997243106365, + 0.009969351813197136, + 0.004333876073360443, + 0.0060190134681761265, + 0.005339097231626511, + 0.004887979477643967, + 0.004269659053534269, + 0.0036829125601798296, + 0.006973040755838156, + 0.006476810667663813, + 0.005188578739762306, + 0.003923734650015831, + 0.004758410155773163, + 0.006703751161694527, + 0.0033993376418948174, + 0.0047895764000713825, + 0.005136054940521717, + 0.00496291508898139, + 0.003069640137255192, + 0.004100833088159561, + 0.007414480205625296, + 0.007369412574917078, + 0.006170003209263086, + 0.0062560513615608215, + 0.006077684927731752, + 0.007456607185304165, + 0.0067326463758945465, + 0.004833046346902847, + 0.01511040236800909, + 0.02116703987121582, + 0.009573535062372684, + 0.016787391155958176, + 0.010407041758298874, + 0.024754831567406654, + 0.014228410087525845, + 0.009918918833136559, + 0.011770800687372684, + 0.016006793826818466, + 0.010990637354552746, + 0.00991507526487112, + 0.013660416938364506, + 0.012191366404294968, + 0.007501576095819473, + 0.011558949016034603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.08901481330394745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068249940872192 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068249940872192, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068249940872192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05455596372485161, + 0.07659398764371872, + 0.07590626925230026, + 0.041471246629953384, + 0.13984167575836182, + 0.12205909937620163, + 0.021505480632185936, + 0.01986819878220558, + 0.19028590619564056, + 0.25791215896606445 + ], + "credit_map": [ + 0.006844860035926104, + 0.0077375369146466255, + 0.0066738612949848175, + 0.006537022069096565, + 0.007677868939936161, + 0.009350608102977276, + 0.0049537778832018375, + 0.004780429415404797, + 0.009865599684417248, + 0.011617083102464676, + 0.008156044408679008, + 0.009878093376755714, + 0.00866101123392582, + 0.012700275518000126, + 0.005346757359802723, + 0.010369124822318554, + 0.011171539314091206, + 0.005753099452704191, + 0.007510026451200247, + 0.009801323525607586, + 0.007060482632368803, + 0.018647272139787674, + 0.006336330436170101, + 0.009626198559999466, + 0.005040344782173634, + 0.00460481783375144, + 0.005704860202968121, + 0.004563844762742519, + 0.003726092167198658, + 0.008875478990375996, + 0.0038833601865917444, + 0.005072448868304491, + 0.02404351532459259, + 0.01481631863862276, + 0.013799735344946384, + 0.009218999184668064, + 0.015502339228987694, + 0.03219911456108093, + 0.012257909402251244, + 0.01800374872982502, + 0.02069697715342045, + 0.012156275101006031, + 0.01315420214086771, + 0.009775885380804539, + 0.012211247347295284, + 0.025754081085324287, + 0.011301414109766483, + 0.01700901798903942, + 0.002941930666565895, + 0.0029597440734505653, + 0.0019757570698857307, + 0.0016883910866454244, + 0.0028701070696115494, + 0.004973541013896465, + 0.0014598221750929952, + 0.0026361881755292416, + 0.0031355496030300856, + 0.0031206891871988773, + 0.001989366952329874, + 0.002372381743043661, + 0.002729382598772645, + 0.0031337530817836523, + 0.0019128115382045507, + 0.0014742636121809483, + 0.023830994963645935, + 0.01697121001780033, + 0.02861850894987583, + 0.0411108061671257, + 0.014341259375214577, + 0.03093479387462139, + 0.022560512647032738, + 0.011917821131646633, + 0.0331936739385128, + 0.03906923905014992, + 0.047044966369867325, + 0.02686985768377781, + 0.031915437430143356, + 0.03219164162874222, + 0.026530029252171516, + 0.021097319200634956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.016068249940872192 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly Time Lord wearing a crimson velvet robe with gold embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the character wear a massive, rigid, fan-shaped collar made of polished metal and dark fabric?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character's right hand extended with the index finger pointing firmly at something off-screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a domed council chamber featuring curved wooden benches and huge stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a beam of soft light from a central circular skylight, casting a long shadow behind the character?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033697061240673065 + ], + "opposite_logprob_tokens": [ + -3.4086971282958984 + ], + "logprob_reward": -0.033697061240673065, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033697061240673065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12969055771827698, + 0.05767502635717392, + 0.0646345466375351, + 0.08989019691944122, + 0.11508595198392868, + 0.21231980621814728, + 0.029099056497216225, + 0.03766973316669464, + 0.05417848378419876, + 0.20975670218467712 + ], + "credit_map": [ + 0.015903696417808533, + 0.019232546910643578, + 0.015278035774827003, + 0.010645740665495396, + 0.020307103171944618, + 0.022382060065865517, + 0.01445531751960516, + 0.011486071161925793, + 0.010029611177742481, + 0.005446272436529398, + 0.006061469670385122, + 0.004153282381594181, + 0.007520804647356272, + 0.015874575823545456, + 0.0044044614769518375, + 0.004184552002698183, + 0.011987054720520973, + 0.005405626259744167, + 0.007532121613621712, + 0.006665636785328388, + 0.006698534358292818, + 0.011878092773258686, + 0.00716543011367321, + 0.007302050944417715, + 0.008691012859344482, + 0.008141839876770973, + 0.011629248037934303, + 0.010006671771407127, + 0.01176453847438097, + 0.015540028922259808, + 0.013170073740184307, + 0.010946787893772125, + 0.020488375797867775, + 0.011744407936930656, + 0.010870804078876972, + 0.0071504139341413975, + 0.012822120450437069, + 0.029271617531776428, + 0.010494359768927097, + 0.012243852019309998, + 0.0378539115190506, + 0.02338084578514099, + 0.01910949870944023, + 0.014722111634910107, + 0.02555960975587368, + 0.04967721179127693, + 0.021158821880817413, + 0.020857790485024452, + 0.0037909746170043945, + 0.004818385001271963, + 0.0029418538324534893, + 0.003540291218087077, + 0.0033747332636266947, + 0.003703602822497487, + 0.003453704994171858, + 0.003475511446595192, + 0.005867824889719486, + 0.004801921546459198, + 0.0031815513502806425, + 0.0036282918881624937, + 0.005807379260659218, + 0.0065687475726008415, + 0.004872215446084738, + 0.0029418005142360926, + 0.005267962347716093, + 0.007234275806695223, + 0.007757349405437708, + 0.011374438181519508, + 0.006067527923732996, + 0.007138807792216539, + 0.006576021201908588, + 0.0027621006593108177, + 0.03668776527047157, + 0.029248161241412163, + 0.034589655697345734, + 0.014156480319797993, + 0.01510810200124979, + 0.049163639545440674, + 0.011802379041910172, + 0.019000519067049026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.033697061240673065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_6/index_tensor([690, 690])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_6/index_tensor([690, 690])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fa9ab6c2ed49a6696d6bef46bbc3878223da6ed8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_6/index_tensor([690, 690])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly tilted, overhead angle focuses on a sheet of slightly yellowed rice paper with a natural fiber texture. Two dragonflies, rendered in traditional Chinese ink painting techniques, come to life. One dragonfly, positioned at the upper edge of the frame, its slender body sketched in a single stroke of ink that gradually fades from thick to light, its head dotted with rich, moist ink, lending it a spirited presence. The other, following closely behind, tilted slightly, creates a vivid sense of pursuit. Most striking are their gauzy, almost transparent wings, created with light strokes of ink or almost water. Only a few precise, powerful, and fine lines of ink outline the structure of the wing veins, perfectly conveying the visual effect of the wings' rapid vibrations. The ink smudges naturally across the rice paper, creating a captivating, irregular circle of moisture, particularly at the edge where the dragonfly's body meets the paper. The background is largely white, with only a few touches of pale cyan surrounding the dragonfly, suggesting a morning mist or the moist air of a pond. The entire painting is bathed in a soft, evenly diffused light, devoid of any distinct shadows. The dominant color palette is classic black, white, and gray, creating a tranquil and elegant atmosphere imbued with the Zen and ethereal qualities of classical Eastern aesthetics.", + "seed": 463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_6/index_690_seed_463_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04680572804063558, + "vlm_logprob": -0.04680572804063558, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.04680572804063558, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01571532152593136 + ], + "opposite_logprob_tokens": [ + -4.5157151222229 + ], + "logprob_reward": -0.01571532152593136, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01571532152593136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11620837450027466, + 0.03977879881858826, + 0.5222955346107483, + 0.03789539635181427, + 0.04499475657939911, + 0.01990334689617157, + 0.04351899400353432, + 0.06815534085035324, + 0.085858054459095, + 0.02139141783118248 + ], + "credit_map": [ + 0.01260861661285162, + 0.010725913569331169, + 0.011952624656260014, + 0.009136621840298176, + 0.019200904294848442, + 0.02394384890794754, + 0.019500812515616417, + 0.009139033034443855, + 0.0028935165610164404, + 0.0040143211372196674, + 0.00477052154019475, + 0.003628587583079934, + 0.005904298275709152, + 0.007874894887208939, + 0.0040198122151196, + 0.006672847084701061, + 0.04069947078824043, + 0.06481471657752991, + 0.08058858662843704, + 0.03637218102812767, + 0.06479137390851974, + 0.14187757670879364, + 0.045604243874549866, + 0.04754732921719551, + 0.0030781179666519165, + 0.003779031103476882, + 0.004174376372247934, + 0.0040834397077560425, + 0.004304864909499884, + 0.007015523500740528, + 0.005073792766779661, + 0.00638625118881464, + 0.0032250769436359406, + 0.009348811581730843, + 0.002858264371752739, + 0.005080067086964846, + 0.005647891666740179, + 0.007820687256753445, + 0.0058083003386855125, + 0.00520565640181303, + 0.0012695968616753817, + 0.004130213055759668, + 0.0018305341945961118, + 0.002546166768297553, + 0.0021404584404081106, + 0.003045304911211133, + 0.0026496103964746, + 0.002291461918503046, + 0.004297688603401184, + 0.006372689735144377, + 0.0045732310973107815, + 0.004579405300319195, + 0.005216004326939583, + 0.007036147173494101, + 0.004177335649728775, + 0.007266494445502758, + 0.00960286520421505, + 0.008181321434676647, + 0.008467676118016243, + 0.008372635580599308, + 0.007497616112232208, + 0.009302382357418537, + 0.007472848519682884, + 0.00925799272954464, + 0.00968464370816946, + 0.013564224354922771, + 0.006427098531275988, + 0.007242157589644194, + 0.009474792517721653, + 0.015267658047378063, + 0.006429377477616072, + 0.017768099904060364, + 0.0032978812232613564, + 0.0023169065825641155, + 0.001558747491799295, + 0.002143826335668564, + 0.0022112277802079916, + 0.003442777320742607, + 0.00214915513060987, + 0.004270895849913359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.01571532152593136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329925447702408 + ], + "opposite_logprob_tokens": [ + -3.773299217224121 + ], + "logprob_reward": -0.02329925447702408, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329925447702408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1691693812608719, + 0.055554453283548355, + 0.1992006152868271, + 0.08696889877319336, + 0.09911671280860901, + 0.10878273099660873, + 0.02826952189207077, + 0.11646329611539841, + 0.09924276173114777, + 0.03723169490695 + ], + "credit_map": [ + 0.01882249303162098, + 0.01702439971268177, + 0.019961455836892128, + 0.01319333165884018, + 0.02924046665430069, + 0.02890782244503498, + 0.02731999196112156, + 0.01469939760863781, + 0.005243674851953983, + 0.006183624733239412, + 0.0064903562888503075, + 0.00798315554857254, + 0.007783057633787394, + 0.010555434972047806, + 0.005863832775503397, + 0.005451313219964504, + 0.01375050563365221, + 0.0061894366517663, + 0.04311748966574669, + 0.00797274149954319, + 0.027276527136564255, + 0.07543065398931503, + 0.015097538009285927, + 0.010365713387727737, + 0.008951985277235508, + 0.010898365639150143, + 0.009313101880252361, + 0.00843783374875784, + 0.010283466428518295, + 0.01800062507390976, + 0.009910338558256626, + 0.011173178441822529, + 0.00792175717651844, + 0.02342715486884117, + 0.007895293645560741, + 0.011421171016991138, + 0.01320505142211914, + 0.016405118629336357, + 0.010162478312849998, + 0.00867867935448885, + 0.007822038605809212, + 0.021067120134830475, + 0.014165268279612064, + 0.011561170220375061, + 0.0146085349842906, + 0.01501820981502533, + 0.011913053691387177, + 0.012627322226762772, + 0.0034833902027457952, + 0.0038195925299078226, + 0.0037431898526847363, + 0.0031005560886114836, + 0.0032735378481447697, + 0.004131839144974947, + 0.002900940366089344, + 0.00381647190079093, + 0.014051469042897224, + 0.01636502891778946, + 0.015230139717459679, + 0.013082163408398628, + 0.015544318594038486, + 0.016075067222118378, + 0.011679229326546192, + 0.01443586964160204, + 0.00953559298068285, + 0.014104836620390415, + 0.008386909030377865, + 0.009123855270445347, + 0.00955806765705347, + 0.016392383724451065, + 0.008954795077443123, + 0.02318630926311016, + 0.004357798956334591, + 0.005006797611713409, + 0.0034687744919210672, + 0.004419384989887476, + 0.004700066521763802, + 0.006009224336594343, + 0.0040006437338888645, + 0.0052690026350319386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.02329925447702408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071676388382912 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071676388382912, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071676388382912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2104090005159378, + 0.05403212830424309, + 0.1638534665107727, + 0.17905418574810028, + 0.042056385427713394, + 0.02420271746814251, + 0.03325183317065239, + 0.19009709358215332, + 0.08058436959981918, + 0.022458789870142937 + ], + "credit_map": [ + 0.020862141624093056, + 0.02751680091023445, + 0.01953883096575737, + 0.015234441496431828, + 0.03940517455339432, + 0.04285096004605293, + 0.03051544353365898, + 0.014485218562185764, + 0.0039019642863422632, + 0.004058739636093378, + 0.007044797297567129, + 0.00420449860394001, + 0.0075890077278018, + 0.012453878298401833, + 0.005701563786715269, + 0.009077679365873337, + 0.016639530658721924, + 0.013489861972630024, + 0.02047945372760296, + 0.013151289895176888, + 0.02168554998934269, + 0.03986729308962822, + 0.0189578328281641, + 0.019582662731409073, + 0.017670566216111183, + 0.027155712246894836, + 0.016851598396897316, + 0.02662229724228382, + 0.02412700280547142, + 0.030186522752046585, + 0.01893024519085884, + 0.017510242760181427, + 0.004309268668293953, + 0.00710650347173214, + 0.004186874721199274, + 0.004462788347154856, + 0.004754798952490091, + 0.006926249712705612, + 0.005728760734200478, + 0.004581141751259565, + 0.0020688327495008707, + 0.003683358896523714, + 0.0022598891519010067, + 0.002400398952886462, + 0.0032636679243296385, + 0.0035341354086995125, + 0.003483155509456992, + 0.0035092781763523817, + 0.0036284958478063345, + 0.003996394574642181, + 0.0038891099393367767, + 0.0038112374022603035, + 0.003641746938228607, + 0.004831999074667692, + 0.0041705528274178505, + 0.005282295402139425, + 0.025377357378602028, + 0.022434592247009277, + 0.02254071645438671, + 0.019153907895088196, + 0.02131974883377552, + 0.027970170602202415, + 0.018644876778125763, + 0.03265572711825371, + 0.007930045947432518, + 0.010951753705739975, + 0.005800133105367422, + 0.007152063772082329, + 0.009161883033812046, + 0.014589725062251091, + 0.0060936217196285725, + 0.018905142322182655, + 0.0031624706462025642, + 0.002638740697875619, + 0.001864967169240117, + 0.0026430608704686165, + 0.0023510665632784367, + 0.0035922094248235226, + 0.002459468087181449, + 0.003746805479750037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.011071676388382912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090870730578899 + ], + "opposite_logprob_tokens": [ + -5.504090785980225 + ], + "logprob_reward": -0.004090870730578899, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090870730578899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1451496183872223, + 0.06768009811639786, + 0.3234376907348633, + 0.07700050622224808, + 0.04651479050517082, + 0.04266523942351341, + 0.0339570976793766, + 0.11030595749616623, + 0.12470046430826187, + 0.028588512912392616 + ], + "credit_map": [ + 0.017612507566809654, + 0.010239137336611748, + 0.017650596797466278, + 0.011028464883565903, + 0.021267730742692947, + 0.030974378809332848, + 0.023981260135769844, + 0.012395555153489113, + 0.004859126638621092, + 0.010067911818623543, + 0.006516993045806885, + 0.006721005775034428, + 0.011148684658110142, + 0.011566895060241222, + 0.009829394519329071, + 0.006970091722905636, + 0.018181269988417625, + 0.01888119988143444, + 0.06602108478546143, + 0.013553549535572529, + 0.04265638440847397, + 0.12346602976322174, + 0.026198184117674828, + 0.014480043202638626, + 0.005609181709587574, + 0.009918635711073875, + 0.00777604291215539, + 0.006949492264539003, + 0.009401153773069382, + 0.014704885892570019, + 0.010690159164369106, + 0.011950964108109474, + 0.0049188463017344475, + 0.007382330019026995, + 0.004118865821510553, + 0.005942157004028559, + 0.006444136146456003, + 0.007188889198005199, + 0.0051857102662324905, + 0.0053338585421442986, + 0.003134810132905841, + 0.008927780203521252, + 0.003953131381422281, + 0.004628099966794252, + 0.004693822469562292, + 0.00667828693985939, + 0.005316645372658968, + 0.0053326706402003765, + 0.003748724004253745, + 0.004085905384272337, + 0.0036129660438746214, + 0.003859788179397583, + 0.003334721550345421, + 0.005022130440920591, + 0.004814846906810999, + 0.00547802122309804, + 0.014854432083666325, + 0.013425772078335285, + 0.014338696375489235, + 0.013884350657463074, + 0.01320973876863718, + 0.01470900233834982, + 0.011329510249197483, + 0.01455447357147932, + 0.012849885039031506, + 0.019312754273414612, + 0.009115329012274742, + 0.01063639298081398, + 0.014252020977437496, + 0.021904705092310905, + 0.009346791543066502, + 0.027282584458589554, + 0.003356456756591797, + 0.003249663859605789, + 0.002600804204121232, + 0.0034782446455210447, + 0.00369673827663064, + 0.0047396449372172356, + 0.0032371277920901775, + 0.004229834768921137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.004090870730578899 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985151708126068 + ], + "opposite_logprob_tokens": [ + -1.8048515319824219 + ], + "logprob_reward": -0.17985151708126068, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985151708126068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13064417243003845, + 0.0599718801677227, + 0.45096302032470703, + 0.052060317248106, + 0.029311785474419594, + 0.034499458968639374, + 0.02752358466386795, + 0.10487215220928192, + 0.09226979315280914, + 0.017883826047182083 + ], + "credit_map": [ + 0.01543968915939331, + 0.013801509514451027, + 0.013496404513716698, + 0.008685017935931683, + 0.021406961604952812, + 0.0272494126111269, + 0.019606707617640495, + 0.010958466678857803, + 0.005627714097499847, + 0.005707424134016037, + 0.0067626661621034145, + 0.005282439757138491, + 0.009460389614105225, + 0.009005985222756863, + 0.008589202538132668, + 0.00953605491667986, + 0.03438689932227135, + 0.06307422369718552, + 0.06532418727874756, + 0.03387438505887985, + 0.0546448677778244, + 0.1189170628786087, + 0.039865970611572266, + 0.040875449776649475, + 0.00391389150172472, + 0.005318284500390291, + 0.005982672795653343, + 0.0049283187836408615, + 0.006192790810018778, + 0.009761353023350239, + 0.007028899621218443, + 0.008934106677770615, + 0.0030312060844153166, + 0.0035555658396333456, + 0.0028608161956071854, + 0.003196865087375045, + 0.0027102665044367313, + 0.004725626204162836, + 0.0051522464491426945, + 0.004079192411154509, + 0.002791033824905753, + 0.007981781847774982, + 0.0036761239171028137, + 0.003895430825650692, + 0.0047133928164839745, + 0.005285817198455334, + 0.002903357846662402, + 0.003252520225942135, + 0.003891910193488002, + 0.003277513198554516, + 0.002613196847960353, + 0.0030856491066515446, + 0.002519267611205578, + 0.004176134709268808, + 0.003927450627088547, + 0.004032463300973177, + 0.014869992621243, + 0.011762047186493874, + 0.012324991635978222, + 0.011358117684721947, + 0.010651513002812862, + 0.01518543902784586, + 0.010639877058565617, + 0.018080178648233414, + 0.010414382442831993, + 0.013483138754963875, + 0.006997816730290651, + 0.007099769078195095, + 0.01064614299684763, + 0.016976187005639076, + 0.0072510624304413795, + 0.01940128766000271, + 0.0019037221791222692, + 0.002167289610952139, + 0.0013902655337005854, + 0.0022408219520002604, + 0.0021527702920138836, + 0.0027417398523539305, + 0.0021071135997772217, + 0.003180102910846472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.17985151708126068 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly tilted, overhead angle focuses on a sheet of slightly yellowed rice paper with a natural fiber texture. Two dragonflies, rendered in traditional Chinese ink painting techniques, come to life. One dragonfly, positioned at the upper edge of the frame, its slender body sketched in a single stroke of ink that gradually fades from thick to light, its head dotted with rich, moist ink, lending it a spirited presence. The other, following closely behind, tilted slightly, creates a vivid sense of pursuit. Most striking are their gauzy, almost transparent wings, created with light strokes of ink or almost water. Only a few precise, powerful, and fine lines of ink outline the structure of the wing veins, perfectly conveying the visual effect of the wings' rapid vibrations. The ink smudges naturally across the rice paper, creating a captivating, irregular circle of moisture, particularly at the edge where the dragonfly's body meets the paper. The background is largely white, with only a few touches of pale cyan surrounding the dragonfly, suggesting a morning mist or the moist air of a pond. The entire painting is bathed in a soft, evenly diffused light, devoid of any distinct shadows. The dominant color palette is classic black, white, and gray, creating a tranquil and elegant atmosphere imbued with the Zen and ethereal qualities of classical Eastern aesthetics.", + "seed": 464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_6/index_690_seed_464_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.014119898620992898, + "vlm_logprob": -0.014119898620992898, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.014119898620992898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015397603623569012 + ], + "opposite_logprob_tokens": [ + -5.015397548675537 + ], + "logprob_reward": -0.015397603623569012, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.015397603623569012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09093664586544037, + 0.1825406700372696, + 0.03355949744582176, + 0.045197390019893646, + 0.03812861442565918, + 0.03241746872663498, + 0.2725391089916229, + 0.1623559445142746, + 0.06769044697284698, + 0.07463424652814865 + ], + "credit_map": [ + 0.00893407966941595, + 0.012747593224048615, + 0.01288395095616579, + 0.007104336749762297, + 0.007179831620305777, + 0.02205633372068405, + 0.008708136156201363, + 0.01132237445563078, + 0.01948866806924343, + 0.02283884957432747, + 0.026177575811743736, + 0.01951432228088379, + 0.020286137238144875, + 0.032755207270383835, + 0.016192497685551643, + 0.025287386029958725, + 0.003533587558194995, + 0.004370482638478279, + 0.00473117595538497, + 0.004640831612050533, + 0.003638811642304063, + 0.005182851571589708, + 0.0026426713448017836, + 0.004819084424525499, + 0.003924517892301083, + 0.00587098253890872, + 0.0047402470372617245, + 0.0073315538465976715, + 0.004254218190908432, + 0.006825735326856375, + 0.005344889592379332, + 0.006905245129019022, + 0.0037661725655198097, + 0.00740337697789073, + 0.0038297574501484632, + 0.004770947154611349, + 0.005568454507738352, + 0.004589884541928768, + 0.004269106313586235, + 0.0039309109561145306, + 0.00299466704018414, + 0.00530899316072464, + 0.0035307412035763264, + 0.0034425046760588884, + 0.004269183147698641, + 0.0036244953516870737, + 0.003861819626763463, + 0.00538506219163537, + 0.019341295585036278, + 0.017059218138456345, + 0.05773141607642174, + 0.022863585501909256, + 0.01816151849925518, + 0.07032379508018494, + 0.029745057225227356, + 0.03731323778629303, + 0.008514404296875, + 0.011901295743882656, + 0.04720509424805641, + 0.01079291757196188, + 0.010313479229807854, + 0.051754552870988846, + 0.011382061056792736, + 0.010492113418877125, + 0.007204866036772728, + 0.012773150578141212, + 0.0075102332048118114, + 0.009208142757415771, + 0.009141205810010433, + 0.008207364939153194, + 0.005453299731016159, + 0.008192180655896664, + 0.007973789237439632, + 0.009678652510046959, + 0.009954853914678097, + 0.00930323451757431, + 0.011574085801839828, + 0.00939928274601698, + 0.007068302482366562, + 0.00968204252421856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.015397603623569012 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029823778197169304 + ], + "opposite_logprob_tokens": [ + -3.5298237800598145 + ], + "logprob_reward": -0.029823778197169304, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029823778197169304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06439292430877686, + 0.02892719954252243, + 0.04557066783308983, + 0.029503803700208664, + 0.041964367032051086, + 0.06903423368930817, + 0.13831940293312073, + 0.16626952588558197, + 0.29211878776550293, + 0.12389902770519257 + ], + "credit_map": [ + 0.008360055275261402, + 0.011667797341942787, + 0.00787219125777483, + 0.006764543708413839, + 0.00828989502042532, + 0.006278455723077059, + 0.00772427674382925, + 0.007435705978423357, + 0.00280900951474905, + 0.004239807371050119, + 0.003967577591538429, + 0.0034128346014767885, + 0.003181483130902052, + 0.0049024117179214954, + 0.0029764471109956503, + 0.0034376289695501328, + 0.003705743933096528, + 0.00665967958047986, + 0.005396632477641106, + 0.007700604852288961, + 0.005346819292753935, + 0.0060194008983671665, + 0.005107759032398462, + 0.005634028930217028, + 0.0025393327232450247, + 0.004553420003503561, + 0.003295404138043523, + 0.005137170664966106, + 0.0038860775530338287, + 0.002827271819114685, + 0.003997383173555136, + 0.003267743857577443, + 0.004144473932683468, + 0.00795825570821762, + 0.004047389142215252, + 0.005742216948419809, + 0.004968519788235426, + 0.005124351009726524, + 0.005543031729757786, + 0.0044361287727952, + 0.005974280182272196, + 0.01049739308655262, + 0.00864490121603012, + 0.010400024242699146, + 0.008619794622063637, + 0.00705297477543354, + 0.00990132987499237, + 0.007943537086248398, + 0.010188083164393902, + 0.01052846573293209, + 0.03172733634710312, + 0.01471503172069788, + 0.010558010078966618, + 0.022837374359369278, + 0.0210864394903183, + 0.016678672283887863, + 0.011788839474320412, + 0.017122050747275352, + 0.018141407519578934, + 0.04500335827469826, + 0.016376160085201263, + 0.02264462411403656, + 0.017949827015399933, + 0.017243247479200363, + 0.018977435305714607, + 0.03043714165687561, + 0.057922475039958954, + 0.035096149891614914, + 0.025693019852042198, + 0.07324624806642532, + 0.01901782862842083, + 0.03172849118709564, + 0.009733862243592739, + 0.019636863842606544, + 0.01349217165261507, + 0.02109532430768013, + 0.016202453523874283, + 0.018006030470132828, + 0.011352273635566235, + 0.014380043372511864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.029823778197169304 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583242177963257 + ], + "opposite_logprob_tokens": [ + -3.895583152770996 + ], + "logprob_reward": -0.020583242177963257, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583242177963257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07509128749370575, + 0.07788868993520737, + 0.03903064504265785, + 0.04505305737257004, + 0.14955690503120422, + 0.10755324363708496, + 0.1234055608510971, + 0.15100356936454773, + 0.07745268940925598, + 0.15396437048912048 + ], + "credit_map": [ + 0.01029497105628252, + 0.01098426803946495, + 0.01006860751658678, + 0.0075941141694784164, + 0.008647704496979713, + 0.010011040605604649, + 0.0077495817095041275, + 0.009740998037159443, + 0.0072642965242266655, + 0.011794377118349075, + 0.011344106867909431, + 0.007623704615980387, + 0.011169390752911568, + 0.011435149237513542, + 0.006908481474965811, + 0.010349180549383163, + 0.004053010605275631, + 0.007002708967775106, + 0.004472318571060896, + 0.006956320721656084, + 0.004294087179005146, + 0.0035374595317989588, + 0.003979439847171307, + 0.0047352975234389305, + 0.004003632348030806, + 0.009879636578261852, + 0.00470650102943182, + 0.006449433974921703, + 0.007107389625161886, + 0.004053750075399876, + 0.0036476922687143087, + 0.005205017514526844, + 0.014032114297151566, + 0.019338157027959824, + 0.02012067474424839, + 0.013133923523128033, + 0.02042384445667267, + 0.02141479216516018, + 0.016324074938893318, + 0.024769313633441925, + 0.009918018244206905, + 0.020030483603477478, + 0.014029636979103088, + 0.012840509414672852, + 0.01606234721839428, + 0.009833219461143017, + 0.009701240807771683, + 0.01513778604567051, + 0.00997863244265318, + 0.012596502900123596, + 0.01375605259090662, + 0.011804429814219475, + 0.01044486928731203, + 0.02544524148106575, + 0.012765427120029926, + 0.026614399626851082, + 0.008497126400470734, + 0.010074717923998833, + 0.03631063550710678, + 0.02777620404958725, + 0.009900783188641071, + 0.03362318128347397, + 0.012437666766345501, + 0.012383252382278442, + 0.006439519114792347, + 0.009487488307058811, + 0.009021195583045483, + 0.014356901869177818, + 0.011631296947598457, + 0.008393028751015663, + 0.00746895233169198, + 0.010654307901859283, + 0.012776879593729973, + 0.03374183923006058, + 0.017067326232790947, + 0.025093739852309227, + 0.025653427466750145, + 0.01840515434741974, + 0.009650220163166523, + 0.011575784534215927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.020583242177963257 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011797142215073109 + ], + "opposite_logprob_tokens": [ + -6.7511796951293945 + ], + "logprob_reward": -0.0011797142215073109, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011797142215073109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08050312846899033, + 0.08382365852594376, + 0.050188373774290085, + 0.04483899474143982, + 0.02925219014286995, + 0.05708952248096466, + 0.27064892649650574, + 0.21209564805030823, + 0.08154784142971039, + 0.09001167863607407 + ], + "credit_map": [ + 0.009659274481236935, + 0.014858291484415531, + 0.010578293353319168, + 0.0071981362998485565, + 0.010996678844094276, + 0.009830040857195854, + 0.008286661468446255, + 0.009095740504562855, + 0.007864557206630707, + 0.011027304455637932, + 0.012896223925054073, + 0.008158120326697826, + 0.009059303440153599, + 0.015550704672932625, + 0.006833122577518225, + 0.012434322386980057, + 0.005140590015798807, + 0.007255257572978735, + 0.006184219848364592, + 0.007523656357079744, + 0.0061254678294062614, + 0.007987428456544876, + 0.004474264103919268, + 0.005497486796230078, + 0.004193403758108616, + 0.007120230235159397, + 0.0050880261696875095, + 0.006387031637132168, + 0.0061949677765369415, + 0.005160279106348753, + 0.004276596009731293, + 0.006418460048735142, + 0.00339873181656003, + 0.006135514937341213, + 0.002834669779986143, + 0.0034720911644399166, + 0.0034453445114195347, + 0.003415758488699794, + 0.003568396670743823, + 0.0029816837050020695, + 0.005282622762024403, + 0.008019962348043919, + 0.007654496002942324, + 0.008535336703062057, + 0.00676679564639926, + 0.006359429564327002, + 0.00721724284812808, + 0.007253638468682766, + 0.020550748333334923, + 0.019264699891209602, + 0.049296483397483826, + 0.02474711462855339, + 0.020290544256567955, + 0.06277495622634888, + 0.03290056437253952, + 0.04082382097840309, + 0.013914654962718487, + 0.029576940461993217, + 0.03158401697874069, + 0.047344695776700974, + 0.02021891064941883, + 0.03035784512758255, + 0.022557424381375313, + 0.016541149467229843, + 0.007270990405231714, + 0.01505034975707531, + 0.008645708672702312, + 0.013323628343641758, + 0.011358332820236683, + 0.010134869255125523, + 0.006259449757635593, + 0.00950451847165823, + 0.007192505989223719, + 0.014813820831477642, + 0.012306373566389084, + 0.014805460348725319, + 0.015171345323324203, + 0.012597762979567051, + 0.0057906657457351685, + 0.007333741523325443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0011797142215073109 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036151548847556114 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036151548847556114, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036151548847556114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15185432136058807, + 0.15401805937290192, + 0.05383490398526192, + 0.020490674301981926, + 0.022174851968884468, + 0.03415708616375923, + 0.04571940377354622, + 0.3454498052597046, + 0.09363172948360443, + 0.07866913825273514 + ], + "credit_map": [ + 0.01028643548488617, + 0.021476730704307556, + 0.026290111243724823, + 0.014823040924966335, + 0.01281217485666275, + 0.029864894226193428, + 0.018754230812191963, + 0.01754670776426792, + 0.013435788452625275, + 0.01906747557222843, + 0.02434779889881611, + 0.014810416847467422, + 0.018044643104076385, + 0.028903841972351074, + 0.012086284346878529, + 0.02332182042300701, + 0.005044582299888134, + 0.0056739985011518, + 0.008954528719186783, + 0.00660679442808032, + 0.005581673700362444, + 0.011103027500212193, + 0.004428710788488388, + 0.006441586650907993, + 0.0018581351032480597, + 0.0033881436102092266, + 0.0029750068206340075, + 0.002306625247001648, + 0.0025979874189943075, + 0.0031006711069494486, + 0.0014848018763586879, + 0.0027793031185865402, + 0.0023350233677774668, + 0.0035770926624536514, + 0.002641279948875308, + 0.0021481250878423452, + 0.003447187365964055, + 0.003025442361831665, + 0.0028370472136884928, + 0.0021636548917740583, + 0.0020158032421022654, + 0.005438375286757946, + 0.005185648798942566, + 0.005237425211817026, + 0.0027716022450476885, + 0.0046988967806100845, + 0.003911093343049288, + 0.004898240324109793, + 0.0026929881423711777, + 0.0054295239970088005, + 0.005258252378553152, + 0.006615209858864546, + 0.004903008230030537, + 0.0077424440532922745, + 0.007848858833312988, + 0.005229122005403042, + 0.016081351786851883, + 0.00947270356118679, + 0.12098459154367447, + 0.018384763970971107, + 0.015596345998346806, + 0.13961338996887207, + 0.011098004877567291, + 0.014218667522072792, + 0.005630220286548138, + 0.012216206640005112, + 0.018034612759947777, + 0.011356879957020283, + 0.008595123887062073, + 0.022539207711815834, + 0.004647016525268555, + 0.010612457990646362, + 0.007019506301730871, + 0.00963505543768406, + 0.013989031314849854, + 0.01337416935712099, + 0.010367101058363914, + 0.012819272466003895, + 0.005260230042040348, + 0.00620477506890893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.0036151548847556114 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly tilted, overhead angle focuses on a sheet of slightly yellowed rice paper with a natural fiber texture. Two dragonflies, rendered in traditional Chinese ink painting techniques, come to life. One dragonfly, positioned at the upper edge of the frame, its slender body sketched in a single stroke of ink that gradually fades from thick to light, its head dotted with rich, moist ink, lending it a spirited presence. The other, following closely behind, tilted slightly, creates a vivid sense of pursuit. Most striking are their gauzy, almost transparent wings, created with light strokes of ink or almost water. Only a few precise, powerful, and fine lines of ink outline the structure of the wing veins, perfectly conveying the visual effect of the wings' rapid vibrations. The ink smudges naturally across the rice paper, creating a captivating, irregular circle of moisture, particularly at the edge where the dragonfly's body meets the paper. The background is largely white, with only a few touches of pale cyan surrounding the dragonfly, suggesting a morning mist or the moist air of a pond. The entire painting is bathed in a soft, evenly diffused light, devoid of any distinct shadows. The dominant color palette is classic black, white, and gray, creating a tranquil and elegant atmosphere imbued with the Zen and ethereal qualities of classical Eastern aesthetics.", + "seed": 465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_6/index_690_seed_465_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.009562004916369915, + "vlm_logprob": -0.009562004916369915, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.009562004916369915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007288888096809387 + ], + "opposite_logprob_tokens": [ + -5.757288932800293 + ], + "logprob_reward": -0.007288888096809387, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007288888096809387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1888784021139145, + 0.05533936992287636, + 0.06016146019101143, + 0.06946234405040741, + 0.0852733701467514, + 0.08736098557710648, + 0.07236409932374954, + 0.07542078197002411, + 0.12200447171926498, + 0.1837346851825714 + ], + "credit_map": [ + 0.016940945759415627, + 0.02390693500638008, + 0.0213452335447073, + 0.021786386147141457, + 0.03211420401930809, + 0.023693375289440155, + 0.02197260782122612, + 0.02711871825158596, + 0.006285900250077248, + 0.005963699426501989, + 0.006793790962547064, + 0.009926355443894863, + 0.004752127453684807, + 0.004675801377743483, + 0.006584308110177517, + 0.010357384569942951, + 0.00644068606197834, + 0.00746974628418684, + 0.00848996639251709, + 0.009033074602484703, + 0.005687343422323465, + 0.006490849889814854, + 0.00751725398004055, + 0.009032540023326874, + 0.008558973670005798, + 0.010494318790733814, + 0.007219096180051565, + 0.006583939306437969, + 0.008998573757708073, + 0.008398672565817833, + 0.008956934325397015, + 0.010251837782561779, + 0.010345392860472202, + 0.009664853103458881, + 0.009714926593005657, + 0.015857726335525513, + 0.010209810920059681, + 0.00876435823738575, + 0.0102809127420187, + 0.01043538749217987, + 0.006800900679081678, + 0.013841508887708187, + 0.009421449154615402, + 0.01128732692450285, + 0.011073114350438118, + 0.01136663369834423, + 0.011962798424065113, + 0.01160725299268961, + 0.007940703071653843, + 0.010660452768206596, + 0.00846861768513918, + 0.009457273408770561, + 0.012885315343737602, + 0.007971908897161484, + 0.007699674926698208, + 0.007280152756720781, + 0.007348471321165562, + 0.01053024921566248, + 0.008914976380765438, + 0.012994866818189621, + 0.011641823686659336, + 0.009728767909109592, + 0.007344430778175592, + 0.006917198654264212, + 0.009174368344247341, + 0.016462868079543114, + 0.01204797625541687, + 0.0200673695653677, + 0.01720093935728073, + 0.01234047394245863, + 0.014794791117310524, + 0.019915683194994926, + 0.012755499221384525, + 0.025095883756875992, + 0.019992049783468246, + 0.029992207884788513, + 0.021844184026122093, + 0.016704972833395004, + 0.02381196990609169, + 0.03353792056441307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.007288888096809387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330007031559944 + ], + "opposite_logprob_tokens": [ + -3.7733001708984375 + ], + "logprob_reward": -0.02330007031559944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330007031559944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15443508327007294, + 0.047939348965883255, + 0.05026879906654358, + 0.05304897204041481, + 0.1480027735233307, + 0.07561902701854706, + 0.26452523469924927, + 0.11811672896146774, + 0.03901180997490883, + 0.04903215542435646 + ], + "credit_map": [ + 0.013831432908773422, + 0.015939747914671898, + 0.01818443089723587, + 0.01877577416598797, + 0.01996488869190216, + 0.020247651264071465, + 0.030520765110850334, + 0.016970394179224968, + 0.003688183845952153, + 0.005057834554463625, + 0.0062030707485973835, + 0.007281752303242683, + 0.007377978414297104, + 0.004709402099251747, + 0.008944623172283173, + 0.004676501266658306, + 0.00610016193240881, + 0.006436001975089312, + 0.006299025844782591, + 0.006059609819203615, + 0.00696893036365509, + 0.004836306907236576, + 0.007061513140797615, + 0.006507251411676407, + 0.003969547804445028, + 0.0075763314962387085, + 0.006940803956240416, + 0.005819275975227356, + 0.008394096978008747, + 0.006050694268196821, + 0.00824267603456974, + 0.006055545061826706, + 0.01079853530973196, + 0.02268318459391594, + 0.01338881254196167, + 0.01495207380503416, + 0.0322076678276062, + 0.016573002561926842, + 0.019926926121115685, + 0.01747257448732853, + 0.006516625639051199, + 0.009836403653025627, + 0.008677680976688862, + 0.008838601410388947, + 0.010740558616816998, + 0.008485989645123482, + 0.011166980490088463, + 0.011356188915669918, + 0.019390849396586418, + 0.03250211849808693, + 0.025833014398813248, + 0.03052867203950882, + 0.05928622931241989, + 0.03813473507761955, + 0.036040160804986954, + 0.02280947007238865, + 0.006834221072494984, + 0.01219527330249548, + 0.01174219325184822, + 0.014839758165180683, + 0.019708793610334396, + 0.018219368532299995, + 0.020481634885072708, + 0.014095486141741276, + 0.003227201523259282, + 0.005798040423542261, + 0.004124955739825964, + 0.007189825642853975, + 0.00455113360658288, + 0.004002489615231752, + 0.004810695070773363, + 0.005307470448315144, + 0.0037813393864780664, + 0.004704327788203955, + 0.005996684543788433, + 0.008604305796325207, + 0.006680910941213369, + 0.004215162713080645, + 0.0071346755139529705, + 0.007914749905467033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.02330007031559944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777512401342392 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.009777512401342392, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777512401342392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1556949019432068, + 0.05413565784692764, + 0.1231633722782135, + 0.08403947949409485, + 0.11566636711359024, + 0.08488056808710098, + 0.17714621126651764, + 0.09541688114404678, + 0.07119280844926834, + 0.03866375982761383 + ], + "credit_map": [ + 0.011276019737124443, + 0.01700935885310173, + 0.01747695356607437, + 0.019938314333558083, + 0.018955515697598457, + 0.01838439516723156, + 0.022239653393626213, + 0.03041469305753708, + 0.004958534613251686, + 0.007151458412408829, + 0.006746492814272642, + 0.006532005500048399, + 0.005943573545664549, + 0.007718225009739399, + 0.007373186759650707, + 0.007712177932262421, + 0.016114825382828712, + 0.01551295816898346, + 0.014240353368222713, + 0.014331348240375519, + 0.020029833540320396, + 0.011244433932006359, + 0.016648266464471817, + 0.015041355043649673, + 0.006487512961030006, + 0.01267197448760271, + 0.012523499317467213, + 0.00995924137532711, + 0.013776907697319984, + 0.009861229918897152, + 0.010373220779001713, + 0.00838589109480381, + 0.010229372419416904, + 0.015299518592655659, + 0.012021927163004875, + 0.01382711622864008, + 0.01740673929452896, + 0.015351017005741596, + 0.01734340563416481, + 0.014187270775437355, + 0.005687838885933161, + 0.012278277426958084, + 0.009705166332423687, + 0.011175747029483318, + 0.011005775071680546, + 0.010410505346953869, + 0.012453163042664528, + 0.012164092622697353, + 0.01369998324662447, + 0.02410198375582695, + 0.017325742170214653, + 0.02001826837658882, + 0.0380883626639843, + 0.024013109505176544, + 0.023427022621035576, + 0.01647173799574375, + 0.006744895596057177, + 0.008936701342463493, + 0.009849242866039276, + 0.011405715718865395, + 0.011670945212244987, + 0.013667838647961617, + 0.013502669520676136, + 0.019638875499367714, + 0.0067318882793188095, + 0.011812848970293999, + 0.008354291319847107, + 0.01498470176011324, + 0.005941458977758884, + 0.006791979540139437, + 0.007885625585913658, + 0.008690017275512218, + 0.003535004099830985, + 0.005839766934514046, + 0.004410065244883299, + 0.007392673287540674, + 0.00352409016340971, + 0.0033387192524969578, + 0.004501405172049999, + 0.006122033577412367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.009777512401342392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195806708186865 + ], + "opposite_logprob_tokens": [ + -6.127195835113525 + ], + "logprob_reward": -0.002195806708186865, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195806708186865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2503601908683777, + 0.09851474314928055, + 0.12613226473331451, + 0.04864465072751045, + 0.08077094703912735, + 0.08522049337625504, + 0.05091949924826622, + 0.1019192785024643, + 0.06909030675888062, + 0.0884275957942009 + ], + "credit_map": [ + 0.021731413900852203, + 0.029401037842035294, + 0.029293956235051155, + 0.03450439125299454, + 0.027519823983311653, + 0.028203129768371582, + 0.03429005667567253, + 0.04541638866066933, + 0.009855255484580994, + 0.011098025366663933, + 0.011040730401873589, + 0.015552756376564503, + 0.010499858297407627, + 0.012473068200051785, + 0.011540140025317669, + 0.01645490899682045, + 0.016747836023569107, + 0.014058415777981281, + 0.01561641227453947, + 0.016368364915251732, + 0.01650700718164444, + 0.011032541282474995, + 0.01648467220366001, + 0.0193170178681612, + 0.004123054910451174, + 0.0057454160414636135, + 0.006234975066035986, + 0.0068776593543589115, + 0.006813388783484697, + 0.006234024651348591, + 0.006307817529886961, + 0.006308312993496656, + 0.007543796673417091, + 0.01374743226915598, + 0.008516598492860794, + 0.008958348073065281, + 0.015383156947791576, + 0.008615030907094479, + 0.009936443530023098, + 0.008070138283073902, + 0.005669266916811466, + 0.013822712004184723, + 0.008411903865635395, + 0.007756299804896116, + 0.013545622117817402, + 0.0121883200481534, + 0.014429081231355667, + 0.009397286921739578, + 0.004324330482631922, + 0.0062859379686415195, + 0.005218245554715395, + 0.0061227306723594666, + 0.009868173860013485, + 0.0065731448121368885, + 0.006637656129896641, + 0.005889279767870903, + 0.009482421912252903, + 0.010348931886255741, + 0.012364218942821026, + 0.012369339354336262, + 0.008335752412676811, + 0.011617138981819153, + 0.011116962879896164, + 0.026284512132406235, + 0.006079921033233404, + 0.009304202161729336, + 0.007354759145528078, + 0.01194716151803732, + 0.006984379608184099, + 0.007043866906315088, + 0.008175341412425041, + 0.012200668454170227, + 0.00683663971722126, + 0.011303894221782684, + 0.008494913578033447, + 0.010945089161396027, + 0.008201020769774914, + 0.009605009108781815, + 0.008945305831730366, + 0.024095723405480385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.002195806708186865 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052477470599114895 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.0052477470599114895, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052477470599114895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22657151520252228, + 0.14779993891716003, + 0.15567243099212646, + 0.046241261065006256, + 0.14833277463912964, + 0.05800238624215126, + 0.02707771398127079, + 0.033160969614982605, + 0.08914262056350708, + 0.06799842417240143 + ], + "credit_map": [ + 0.0362081415951252, + 0.025637827813625336, + 0.028894122689962387, + 0.02885468304157257, + 0.0176519937813282, + 0.022775188088417053, + 0.024923941120505333, + 0.04162561520934105, + 0.014713740907609463, + 0.01899701915681362, + 0.024490365758538246, + 0.031822629272937775, + 0.011181452311575413, + 0.009784966707229614, + 0.013463763520121574, + 0.023346003144979477, + 0.01839470863342285, + 0.020890671759843826, + 0.022375212982296944, + 0.03957376256585121, + 0.006876803003251553, + 0.009476535953581333, + 0.01586424931883812, + 0.022220483049750328, + 0.004390163812786341, + 0.006296778563410044, + 0.00662548653781414, + 0.010816234163939953, + 0.003295769914984703, + 0.002904278924688697, + 0.004970410373061895, + 0.006942135281860828, + 0.018777281045913696, + 0.02451486885547638, + 0.018473658710718155, + 0.020973270758986473, + 0.0109763378277421, + 0.013912856578826904, + 0.0173353124409914, + 0.02336917631328106, + 0.005267327651381493, + 0.009612079709768295, + 0.007257911376655102, + 0.007611825130879879, + 0.005809117574244738, + 0.005909334868192673, + 0.008016446605324745, + 0.008518342860043049, + 0.0019149509025737643, + 0.003114199498668313, + 0.002814336447045207, + 0.0035676192492246628, + 0.004326390102505684, + 0.0035785972140729427, + 0.004462686367332935, + 0.00329893478192389, + 0.003042041789740324, + 0.004999932367354631, + 0.004286633804440498, + 0.0061787026934325695, + 0.0032864806707948446, + 0.0037605094257742167, + 0.004087354987859726, + 0.003519313409924507, + 0.00932625588029623, + 0.012311378493905067, + 0.01255829632282257, + 0.020782314240932465, + 0.00892687775194645, + 0.008079149760305882, + 0.008987816981971264, + 0.008170531131327152, + 0.008230091072618961, + 0.009523854590952396, + 0.010485301725566387, + 0.01670914515852928, + 0.0056807114742696285, + 0.004988224245607853, + 0.00685814069584012, + 0.005522958468645811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0052477470599114895 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly tilted, overhead angle focuses on a sheet of slightly yellowed rice paper with a natural fiber texture. Two dragonflies, rendered in traditional Chinese ink painting techniques, come to life. One dragonfly, positioned at the upper edge of the frame, its slender body sketched in a single stroke of ink that gradually fades from thick to light, its head dotted with rich, moist ink, lending it a spirited presence. The other, following closely behind, tilted slightly, creates a vivid sense of pursuit. Most striking are their gauzy, almost transparent wings, created with light strokes of ink or almost water. Only a few precise, powerful, and fine lines of ink outline the structure of the wing veins, perfectly conveying the visual effect of the wings' rapid vibrations. The ink smudges naturally across the rice paper, creating a captivating, irregular circle of moisture, particularly at the edge where the dragonfly's body meets the paper. The background is largely white, with only a few touches of pale cyan surrounding the dragonfly, suggesting a morning mist or the moist air of a pond. The entire painting is bathed in a soft, evenly diffused light, devoid of any distinct shadows. The dominant color palette is classic black, white, and gray, creating a tranquil and elegant atmosphere imbued with the Zen and ethereal qualities of classical Eastern aesthetics.", + "seed": 466, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000013/group_6/index_690_seed_466_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011441646795719861, + "vlm_logprob": -0.011441646795719861, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.011441646795719861, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010361921973526478 + ], + "opposite_logprob_tokens": [ + -5.010362148284912 + ], + "logprob_reward": -0.010361921973526478, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010361921973526478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6581552624702454, + 0.029511911794543266, + 0.0225997231900692, + 0.05097779259085655, + 0.06647691875696182, + 0.057347387075424194, + 0.028284363448619843, + 0.012407130561769009, + 0.05245571583509445, + 0.02178378403186798 + ], + "credit_map": [ + 0.042059917002916336, + 0.0219362061470747, + 0.2511652112007141, + 0.19002141058444977, + 0.034561675041913986, + 0.055487845093011856, + 0.036203883588314056, + 0.026719072833657265, + 0.0029473486356437206, + 0.0027268449775874615, + 0.003836277639493346, + 0.0037544225342571735, + 0.004021703731268644, + 0.003977276850491762, + 0.003289085580036044, + 0.0049589527770876884, + 0.0021470757201313972, + 0.0027164060156792402, + 0.004435343202203512, + 0.0021318751387298107, + 0.0034214153420180082, + 0.002609197050333023, + 0.0027770784217864275, + 0.0023613320663571358, + 0.004526094999164343, + 0.005993253085762262, + 0.008036615327000618, + 0.005663494113832712, + 0.007095132954418659, + 0.005736310500651598, + 0.007368383463472128, + 0.006558507680892944, + 0.007298330310732126, + 0.0073976218700408936, + 0.010189730674028397, + 0.005342418793588877, + 0.009827693924307823, + 0.008483797311782837, + 0.009076896123588085, + 0.008860427886247635, + 0.0055162375792860985, + 0.005162524525076151, + 0.008784421719610691, + 0.003560468554496765, + 0.012364950031042099, + 0.007952103391289711, + 0.008023044094443321, + 0.005983641371130943, + 0.00412300368770957, + 0.0025415560230612755, + 0.004393996205180883, + 0.0018025990575551987, + 0.004392743576318026, + 0.004284119699150324, + 0.003766446141526103, + 0.002979897428303957, + 0.001175641198642552, + 0.001305697369389236, + 0.0015718169743195176, + 0.0018396335653960705, + 0.001533881644718349, + 0.002556371269747615, + 0.0009671137086115777, + 0.0014569744234904647, + 0.005604236386716366, + 0.004897988401353359, + 0.008146878331899643, + 0.007253843359649181, + 0.006766369100660086, + 0.008065217174589634, + 0.005971136502921581, + 0.00575004518032074, + 0.0020730288233608007, + 0.0021514641121029854, + 0.0030796255450695753, + 0.002413996495306492, + 0.00227418914437294, + 0.0041784001514315605, + 0.0022100855130702257, + 0.0034029947128146887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.010361921973526478 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059842273592949 + ], + "opposite_logprob_tokens": [ + -3.8955984115600586 + ], + "logprob_reward": -0.02059842273592949, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059842273592949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22872605919837952, + 0.1004597544670105, + 0.05512271821498871, + 0.10021819174289703, + 0.03786073625087738, + 0.03822440654039383, + 0.09467941522598267, + 0.10341614484786987, + 0.2070150226354599, + 0.03427760303020477 + ], + "credit_map": [ + 0.015065805055201054, + 0.009961536154150963, + 0.07445728778839111, + 0.051786087453365326, + 0.02314673364162445, + 0.024567419663071632, + 0.015555327758193016, + 0.014185847714543343, + 0.008081230334937572, + 0.008273478597402573, + 0.010986220091581345, + 0.011429276317358017, + 0.01679840125143528, + 0.012779549695551395, + 0.010416275821626186, + 0.021695313975214958, + 0.0059496574103832245, + 0.006885960698127747, + 0.008124034851789474, + 0.00534577714279294, + 0.00671487208455801, + 0.007410443853586912, + 0.007447862531989813, + 0.0072441003285348415, + 0.01180452760308981, + 0.016433605924248695, + 0.014405316673219204, + 0.010915152728557587, + 0.012898124754428864, + 0.008244994096457958, + 0.014816910959780216, + 0.010699542239308357, + 0.003992205951362848, + 0.003466614056378603, + 0.0056148916482925415, + 0.00276631279848516, + 0.006138796918094158, + 0.005126405507326126, + 0.005702401977032423, + 0.005053101573139429, + 0.00411817105486989, + 0.002824328141286969, + 0.005882320459932089, + 0.0019769410137087107, + 0.007406373508274555, + 0.00613265810534358, + 0.0055870371870696545, + 0.0042965710163116455, + 0.008148947730660439, + 0.009357583709061146, + 0.014262458309531212, + 0.0069182380102574825, + 0.013910189270973206, + 0.01530990470200777, + 0.01899341493844986, + 0.00777866505086422, + 0.008641310967504978, + 0.010776247829198837, + 0.013078550808131695, + 0.017884409055113792, + 0.013407890684902668, + 0.0172285009175539, + 0.010184279642999172, + 0.01221494097262621, + 0.021559765562415123, + 0.021191701292991638, + 0.028114041313529015, + 0.03562002256512642, + 0.02430751919746399, + 0.02960868738591671, + 0.027753878384828568, + 0.018859390169382095, + 0.0034869066439568996, + 0.0041705467738211155, + 0.00469077005982399, + 0.003956581931561232, + 0.004205834586173296, + 0.005214720498770475, + 0.002264222363010049, + 0.006288016214966774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.02059842273592949 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052522532641887665 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.0052522532641887665, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052522532641887665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5319502949714661, + 0.0749431923031807, + 0.030208609998226166, + 0.0986684039235115, + 0.020175283774733543, + 0.046904195100069046, + 0.06043101102113724, + 0.05910687893629074, + 0.051444318145513535, + 0.0261678546667099 + ], + "credit_map": [ + 0.034418318420648575, + 0.031503718346357346, + 0.17219439148902893, + 0.12839677929878235, + 0.05704229697585106, + 0.05467332527041435, + 0.026709022000432014, + 0.027012426406145096, + 0.005086095537990332, + 0.006355034653097391, + 0.009658029302954674, + 0.009763271547853947, + 0.011746280826628208, + 0.008572512306272984, + 0.009803013876080513, + 0.01395895704627037, + 0.0029287647921591997, + 0.002967443084344268, + 0.005621833726763725, + 0.0033626705408096313, + 0.00391030078753829, + 0.004024258814752102, + 0.004122690763324499, + 0.0032706488855183125, + 0.012066245079040527, + 0.015674704685807228, + 0.015548450872302055, + 0.008744725026190281, + 0.009595910087227821, + 0.007931282743811607, + 0.01952715404331684, + 0.00957992859184742, + 0.002032669959589839, + 0.0024513197131454945, + 0.00254822033457458, + 0.0016019977629184723, + 0.0033086331095546484, + 0.002557628322392702, + 0.003080969210714102, + 0.0025938458275049925, + 0.004913707729429007, + 0.00407278910279274, + 0.008143211714923382, + 0.00266227382235229, + 0.008413617499172688, + 0.005031932611018419, + 0.006889406591653824, + 0.006777254398912191, + 0.005656513385474682, + 0.004458158276975155, + 0.008469833061099052, + 0.0033016882371157408, + 0.011685595847666264, + 0.01199946179986, + 0.009306415915489197, + 0.005553346127271652, + 0.004899419844150543, + 0.006920981220901012, + 0.007245330605655909, + 0.00995849259197712, + 0.007140723522752523, + 0.010842389427125454, + 0.005720033776015043, + 0.006379508413374424, + 0.00498450780287385, + 0.004191500134766102, + 0.006503504700958729, + 0.007410560268908739, + 0.006347151007503271, + 0.008478725329041481, + 0.008335253223776817, + 0.005193112883716822, + 0.002075296826660633, + 0.0028782235458493233, + 0.003582727164030075, + 0.0029928702861070633, + 0.003214680589735508, + 0.004983698483556509, + 0.002024243585765362, + 0.00441611185669899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0052522532641887665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816759515553713 + ], + "opposite_logprob_tokens": [ + -5.877816677093506 + ], + "logprob_reward": -0.002816759515553713, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816759515553713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3470328748226166, + 0.0514744408428669, + 0.11239656805992126, + 0.09235983341932297, + 0.04626777023077011, + 0.09155689179897308, + 0.11211543530225754, + 0.024494951590895653, + 0.055361587554216385, + 0.06693962961435318 + ], + "credit_map": [ + 0.021077297627925873, + 0.01667110063135624, + 0.11949772387742996, + 0.09297008812427521, + 0.02704920433461666, + 0.03420032933354378, + 0.01831800304353237, + 0.017249101772904396, + 0.005129609722644091, + 0.0051261852495372295, + 0.005449268501251936, + 0.007933126762509346, + 0.006461561657488346, + 0.004391624592244625, + 0.00786848459392786, + 0.009114574640989304, + 0.009983949363231659, + 0.015865188091993332, + 0.013318349607288837, + 0.011357225477695465, + 0.015724876895546913, + 0.013105446472764015, + 0.014173988252878189, + 0.018867531791329384, + 0.007577961776405573, + 0.010378371924161911, + 0.014453263953328133, + 0.011312108486890793, + 0.014787537045776844, + 0.009939776733517647, + 0.015406198799610138, + 0.008504609577357769, + 0.005761029198765755, + 0.004461640492081642, + 0.0061085219494998455, + 0.0029052640311419964, + 0.0069075520150363445, + 0.005619614385068417, + 0.008462178520858288, + 0.006041963119059801, + 0.00931182224303484, + 0.006784607190638781, + 0.015439768321812153, + 0.004999540746212006, + 0.01637527160346508, + 0.012031497433781624, + 0.012909897603094578, + 0.01370447501540184, + 0.010807419195771217, + 0.007318491116166115, + 0.016343610361218452, + 0.006212880834937096, + 0.02014070190489292, + 0.022891122847795486, + 0.018791167065501213, + 0.009610041044652462, + 0.0020139403641223907, + 0.0024131087120622396, + 0.0033522623125463724, + 0.0023009139113128185, + 0.0038826693780720234, + 0.005955628119409084, + 0.0017780319321900606, + 0.0027983940672129393, + 0.006309754215180874, + 0.005102312192320824, + 0.007400037255138159, + 0.00763730239123106, + 0.006553316488862038, + 0.009657848626375198, + 0.007862617261707783, + 0.004838397726416588, + 0.005231567658483982, + 0.0063766431994736195, + 0.008563196286559105, + 0.010329388082027435, + 0.007147714961320162, + 0.013961497694253922, + 0.005569023080170155, + 0.009760595858097076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.002816759515553713 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of a sheet of slightly yellowed rice paper with a natural fiber texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two dragonflies visible in the scene, rendered in traditional Chinese ink painting techniques?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is one dragonfly positioned at the upper edge of the frame with a slender body sketched in a single stroke of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dragonflies have gauzy, almost transparent wings outlined by fine lines of ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle slightly tilted and overhead, with a background largely white and touches of pale cyan?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178876489400864 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178876489400864, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178876489400864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3833733797073364, + 0.10049616545438766, + 0.055525124073028564, + 0.08085920661687851, + 0.013499059714376926, + 0.11585020273923874, + 0.1694854497909546, + 0.024720001965761185, + 0.026061763986945152, + 0.030129635706543922 + ], + "credit_map": [ + 0.022330358624458313, + 0.012941773049533367, + 0.13919946551322937, + 0.11035746335983276, + 0.026906359940767288, + 0.029567822813987732, + 0.023788930848240852, + 0.018281226977705956, + 0.008557946421205997, + 0.00693677319213748, + 0.010348607785999775, + 0.009793201461434364, + 0.015535161830484867, + 0.01420162059366703, + 0.012283861637115479, + 0.022838998585939407, + 0.005295196082442999, + 0.006123764906078577, + 0.008050278760492802, + 0.004153912421315908, + 0.00856721866875887, + 0.008748949505388737, + 0.006128000095486641, + 0.008457805030047894, + 0.00747063010931015, + 0.009976602159440517, + 0.014030778780579567, + 0.008119607344269753, + 0.011300534009933472, + 0.008104631677269936, + 0.01241537556052208, + 0.009441048838198185, + 0.0015162867493927479, + 0.0013226798037067056, + 0.002306051319465041, + 0.0012509507359936833, + 0.0023462979588657618, + 0.0017794921295717359, + 0.0017939575482159853, + 0.0011833440512418747, + 0.010763420723378658, + 0.01060002762824297, + 0.016888586804270744, + 0.00840925332158804, + 0.023773496970534325, + 0.016367889940738678, + 0.01696004532277584, + 0.012087482027709484, + 0.01491408422589302, + 0.012030592188239098, + 0.0304463729262352, + 0.009356926195323467, + 0.029896613210439682, + 0.02738996408879757, + 0.03288682922720909, + 0.01256406307220459, + 0.0019787740893661976, + 0.002116880612447858, + 0.003266253275796771, + 0.00418308237567544, + 0.0032015061005949974, + 0.005196521990001202, + 0.0022569375578314066, + 0.002520044334232807, + 0.0028185993432998657, + 0.0027740043587982655, + 0.003371071768924594, + 0.00424119783565402, + 0.0030888912733644247, + 0.003673449857160449, + 0.0033946912735700607, + 0.0026998594403266907, + 0.002920556580647826, + 0.0034213631879538298, + 0.0041427211835980415, + 0.0050176698714494705, + 0.0032682877499610186, + 0.006298042368143797, + 0.0019327339250594378, + 0.003128261538222432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.018178876489400864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_0/index_tensor([5671, 5671])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_0/index_tensor([5671, 5671])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3979b8805994308569fa6a99aadf1cc04050d67e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_0/index_tensor([5671, 5671])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a low angle, the camera held still, captures a bird foraging on the quiet, moonlit forest floor. Cool moonlight filters through the sparse branches of tall trees, casting dappled shadows on a carpet of dry, curled leaves. The bird's plumage, covered in a complex array of brown, black, and beige feathers, makes it blend almost seamlessly into its surroundings. Its large, round, black eyes gleam with alertness in the dim light, while its extremely long, slender beak is intensely focused on its task. With steady, precise care, it plunges its beak into the pile of dead leaves at its feet, then rhythmically probes up and down, each dip accompanied by the subtle rustling of leaves being gently brushed aside. The moonlight outlines the contours of its back feathers and a silver highlight on its beak. In the background, several moss-covered tree roots meander in deep shadows. The whole scene, with the main colors of dark blue and earthy brown, presents an extremely peaceful, focused and slightly mysterious atmosphere.", + "seed": 5319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_0/index_5671_seed_5319_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10380965936928988, + "vlm_logprob": -0.10380965936928988, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.10380965936928988, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029213471338152885 + ], + "opposite_logprob_tokens": [ + -3.9042134284973145 + ], + "logprob_reward": -0.029213471338152885, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029213471338152885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34234073758125305, + 0.11239548772573471, + 0.097492516040802, + 0.08413399010896683, + 0.025761514902114868, + 0.12156267464160919, + 0.05827046185731888, + 0.03711773827672005, + 0.043978169560432434, + 0.07694672048091888 + ], + "credit_map": [ + 0.030929524451494217, + 0.04340098053216934, + 0.042416058480739594, + 0.027998613193631172, + 0.03172220289707184, + 0.05271989479660988, + 0.06430485844612122, + 0.048848602920770645, + 0.010120362974703312, + 0.01546988170593977, + 0.015006125904619694, + 0.012024215422570705, + 0.009159636683762074, + 0.016643023118376732, + 0.018082963302731514, + 0.015889275819063187, + 0.006845008581876755, + 0.008420764468610287, + 0.013760581612586975, + 0.011064961552619934, + 0.00761910667642951, + 0.015468145720660686, + 0.016563856974244118, + 0.0177500918507576, + 0.007761511020362377, + 0.008151007816195488, + 0.008879981003701687, + 0.00703304773196578, + 0.00785434152930975, + 0.018398728221654892, + 0.011990919709205627, + 0.014064456336200237, + 0.0023429151624441147, + 0.003133936319500208, + 0.003485859837383032, + 0.00272173248231411, + 0.0034521417692303658, + 0.004368789494037628, + 0.0026411290746182203, + 0.0036150114610791206, + 0.008749095723032951, + 0.0069655850529670715, + 0.015271283686161041, + 0.010185318998992443, + 0.012173585593700409, + 0.020817041397094727, + 0.02923048473894596, + 0.018170272931456566, + 0.005424781236797571, + 0.008329738862812519, + 0.007467341143637896, + 0.0066793542355299, + 0.0056689707562327385, + 0.007639066316187382, + 0.008890214376151562, + 0.00817099679261446, + 0.003391249803826213, + 0.005220858845859766, + 0.004608688876032829, + 0.003657486755400896, + 0.003523116232827306, + 0.004407692700624466, + 0.006893965415656567, + 0.0054146782495081425, + 0.0027279318310320377, + 0.003065298078581691, + 0.005288947373628616, + 0.004015149082988501, + 0.005441863555461168, + 0.0046247136779129505, + 0.010590344667434692, + 0.00822392012923956, + 0.004873523488640785, + 0.00539479311555624, + 0.01353104691952467, + 0.014395269565284252, + 0.007908476516604424, + 0.006848854012787342, + 0.0122618917375803, + 0.011732866056263447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.029213471338152885 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743480086326599 + ], + "opposite_logprob_tokens": [ + -0.9743480086326599 + ], + "logprob_reward": -0.4743480086326599, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743480086326599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20818430185317993, + 0.07166974246501923, + 0.06353914737701416, + 0.11298821866512299, + 0.07515785843133926, + 0.10372824966907501, + 0.07730241119861603, + 0.08227920532226562, + 0.1116270050406456, + 0.09352376312017441 + ], + "credit_map": [ + 0.019229033961892128, + 0.023130852729082108, + 0.030284861102700233, + 0.024439509958028793, + 0.020867377519607544, + 0.029057921841740608, + 0.03392596170306206, + 0.027248801663517952, + 0.00670637795701623, + 0.0065921577624976635, + 0.009791525080800056, + 0.007165346760302782, + 0.007233382202684879, + 0.012503987178206444, + 0.01090541947633028, + 0.010771560482680798, + 0.005576263181865215, + 0.008013872429728508, + 0.007364608347415924, + 0.005970647558569908, + 0.005707820411771536, + 0.008603030815720558, + 0.012016856111586094, + 0.010286057367920876, + 0.007104972377419472, + 0.005948924925178289, + 0.01651400700211525, + 0.011407912708818913, + 0.01212810818105936, + 0.013344105333089828, + 0.02698558010160923, + 0.01955462247133255, + 0.00454505393281579, + 0.005951069761067629, + 0.011626990512013435, + 0.005844108294695616, + 0.01112339086830616, + 0.013989000581204891, + 0.01030552014708519, + 0.011772732250392437, + 0.005751506425440311, + 0.006396848242729902, + 0.013291027396917343, + 0.008625050075352192, + 0.01403932273387909, + 0.018350807949900627, + 0.022840749472379684, + 0.014432957395911217, + 0.00562933599576354, + 0.007789088878780603, + 0.014328149147331715, + 0.00766480527818203, + 0.005122323054820299, + 0.012960011139512062, + 0.010548995807766914, + 0.013259713537991047, + 0.004834098741412163, + 0.004093261901289225, + 0.01504991203546524, + 0.005829820409417152, + 0.006405698601156473, + 0.00937544833868742, + 0.02229124866425991, + 0.014399725012481213, + 0.010802767239511013, + 0.008873715065419674, + 0.014230434782803059, + 0.008972645737230778, + 0.01322306040674448, + 0.011643785052001476, + 0.0252997949719429, + 0.01858081854879856, + 0.008739049546420574, + 0.010996953584253788, + 0.011272018775343895, + 0.01364497933536768, + 0.010551219806075096, + 0.011330407112836838, + 0.014530103653669357, + 0.012459038756787777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.4743480086326599 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102861508727074 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.004102861508727074, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102861508727074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.265553742647171, + 0.15692521631717682, + 0.12246189266443253, + 0.07777289301156998, + 0.03918788954615593, + 0.05027029290795326, + 0.12824217975139618, + 0.07278873026371002, + 0.05757926031947136, + 0.029217934235930443 + ], + "credit_map": [ + 0.03097437508404255, + 0.03144797310233116, + 0.046314071863889694, + 0.022953743115067482, + 0.0220868531614542, + 0.030199669301509857, + 0.03597855567932129, + 0.04559849947690964, + 0.019948912784457207, + 0.01781238056719303, + 0.02612808533012867, + 0.014853173866868019, + 0.017230259254574776, + 0.0199901070445776, + 0.017870604991912842, + 0.023091701790690422, + 0.014742979779839516, + 0.014911681413650513, + 0.01723472960293293, + 0.012187298387289047, + 0.010880029760301113, + 0.012639558874070644, + 0.02049359865486622, + 0.019372016191482544, + 0.011976605281233788, + 0.011723221279680729, + 0.009783320128917694, + 0.006555591244250536, + 0.006961718667298555, + 0.007712302729487419, + 0.012781118042767048, + 0.01027901004999876, + 0.003708340460434556, + 0.003727095667272806, + 0.0074555822648108006, + 0.0036266199313104153, + 0.0034559168852865696, + 0.006366850342601538, + 0.004088470246642828, + 0.006759015843272209, + 0.005531552247703075, + 0.005723860580474138, + 0.005261948332190514, + 0.005626467522233725, + 0.006320585031062365, + 0.0070977527648210526, + 0.00832165777683258, + 0.0063864681869745255, + 0.019787978380918503, + 0.02549031935632229, + 0.013657948933541775, + 0.01105690747499466, + 0.011668031103909016, + 0.01350489817559719, + 0.014020559377968311, + 0.019055530428886414, + 0.011751408688724041, + 0.014249610714614391, + 0.008801099844276905, + 0.005292125046253204, + 0.006011068820953369, + 0.006642421241849661, + 0.010143701918423176, + 0.009897293522953987, + 0.005555623210966587, + 0.006780435796827078, + 0.0075194574892520905, + 0.005733325611799955, + 0.006998168770223856, + 0.006759311072528362, + 0.010051828809082508, + 0.008181109093129635, + 0.004764134995639324, + 0.004200414754450321, + 0.004283680580556393, + 0.002975010545924306, + 0.002363205421715975, + 0.003018812043592334, + 0.003461369313299656, + 0.00415130564942956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.004102861508727074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649542272090912 + ], + "opposite_logprob_tokens": [ + -5.379649639129639 + ], + "logprob_reward": -0.004649542272090912, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649542272090912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3363622725009918, + 0.08484888821840286, + 0.10885416716337204, + 0.07394537329673767, + 0.04415779188275337, + 0.15747202932834625, + 0.026594674214720726, + 0.048319824039936066, + 0.0489632673561573, + 0.07048173993825912 + ], + "credit_map": [ + 0.0293462872505188, + 0.028007401153445244, + 0.05438491329550743, + 0.030408533290028572, + 0.027110395953059196, + 0.055918771773576736, + 0.059974305331707, + 0.0512116476893425, + 0.007349164225161076, + 0.005562313366681337, + 0.01757988892495632, + 0.007360916584730148, + 0.008881108835339546, + 0.009596824645996094, + 0.010763630270957947, + 0.01775502972304821, + 0.008149011060595512, + 0.0064142546616494656, + 0.014907291159033775, + 0.011206372641026974, + 0.010493514128029346, + 0.015598188154399395, + 0.021615559235215187, + 0.020469961687922478, + 0.006097240839153528, + 0.005538694094866514, + 0.009040826931595802, + 0.006208155769854784, + 0.006206313148140907, + 0.0077723488211631775, + 0.019616251811385155, + 0.013465533033013344, + 0.005079001188278198, + 0.00685729319229722, + 0.0056203207932412624, + 0.0036427078302949667, + 0.0042031798511743546, + 0.007651523686945438, + 0.005470355972647667, + 0.005633402615785599, + 0.012610970064997673, + 0.01634063571691513, + 0.015061401762068272, + 0.013441022485494614, + 0.01777263917028904, + 0.028437796980142593, + 0.028241125866770744, + 0.025566421449184418, + 0.0028581523802131414, + 0.0025063634384423494, + 0.0050507280975580215, + 0.002899552695453167, + 0.0022040754556655884, + 0.003645152784883976, + 0.0031787226907908916, + 0.004251922480762005, + 0.0034085519146174192, + 0.002482655458152294, + 0.005720496643334627, + 0.003964724950492382, + 0.004870701115578413, + 0.005458167288452387, + 0.014584016986191273, + 0.007830500602722168, + 0.003883956465870142, + 0.0057839336805045605, + 0.007251645904034376, + 0.005444892682135105, + 0.005868867505341768, + 0.005575926974415779, + 0.008034564554691315, + 0.007119473535567522, + 0.005227011628448963, + 0.004907076712697744, + 0.009345624595880508, + 0.01120547205209732, + 0.009228680282831192, + 0.005749627482146025, + 0.01400141604244709, + 0.010816824622452259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.004649542272090912 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734413094818592 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734413094818592, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734413094818592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21614332497119904, + 0.1121358796954155, + 0.22391653060913086, + 0.07620933651924133, + 0.05993244796991348, + 0.05058006942272186, + 0.07884424924850464, + 0.038729920983314514, + 0.08656308799982071, + 0.05694521591067314 + ], + "credit_map": [ + 0.02002406120300293, + 0.021306578069925308, + 0.04183116555213928, + 0.022046204656362534, + 0.016701065003871918, + 0.028002671897411346, + 0.031060215085744858, + 0.03517134115099907, + 0.008966760709881783, + 0.009939861483871937, + 0.02147318609058857, + 0.01056599523872137, + 0.0077017247676849365, + 0.011929034255445004, + 0.018853357061743736, + 0.02270594984292984, + 0.01941656693816185, + 0.01678992062807083, + 0.02847316861152649, + 0.028019489720463753, + 0.01783335953950882, + 0.035290371626615524, + 0.040556080639362335, + 0.03753754869103432, + 0.007250543683767319, + 0.0072647277265787125, + 0.010075908154249191, + 0.009976637549698353, + 0.0060003940016031265, + 0.009417335502803326, + 0.01477767527103424, + 0.011446107178926468, + 0.0035486549604684114, + 0.0037388745695352554, + 0.012221753597259521, + 0.0052307466976344585, + 0.006154767237603664, + 0.009837732650339603, + 0.008373071439564228, + 0.010826848447322845, + 0.004364303313195705, + 0.0033205479849129915, + 0.008903373032808304, + 0.0047910176217556, + 0.004566640593111515, + 0.0069085946306586266, + 0.009170819073915482, + 0.008554765954613686, + 0.006115024909377098, + 0.005110492464154959, + 0.011545129120349884, + 0.008776694536209106, + 0.00638565793633461, + 0.016237810254096985, + 0.013562257401645184, + 0.011111177504062653, + 0.003437332808971405, + 0.0037437197752296925, + 0.00566759193316102, + 0.004049005452543497, + 0.0035547565203160048, + 0.005155707243829966, + 0.00759386969730258, + 0.005527936853468418, + 0.006593591533601284, + 0.006388076115399599, + 0.012230376712977886, + 0.008356256410479546, + 0.01242442149668932, + 0.012413018383085728, + 0.015193901024758816, + 0.012963438406586647, + 0.006841207388788462, + 0.006382634397596121, + 0.007639285176992416, + 0.007080524694174528, + 0.006017496343702078, + 0.006334260106086731, + 0.007935215719044209, + 0.008714588358998299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.006734413094818592 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, the camera held still, captures a bird foraging on the quiet, moonlit forest floor. Cool moonlight filters through the sparse branches of tall trees, casting dappled shadows on a carpet of dry, curled leaves. The bird's plumage, covered in a complex array of brown, black, and beige feathers, makes it blend almost seamlessly into its surroundings. Its large, round, black eyes gleam with alertness in the dim light, while its extremely long, slender beak is intensely focused on its task. With steady, precise care, it plunges its beak into the pile of dead leaves at its feet, then rhythmically probes up and down, each dip accompanied by the subtle rustling of leaves being gently brushed aside. The moonlight outlines the contours of its back feathers and a silver highlight on its beak. In the background, several moss-covered tree roots meander in deep shadows. The whole scene, with the main colors of dark blue and earthy brown, presents an extremely peaceful, focused and slightly mysterious atmosphere.", + "seed": 5320, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_0/index_5671_seed_5320_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2916023432277143, + "vlm_logprob": -0.2916023432277143, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.2916023432277143, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013910956680774689 + ], + "opposite_logprob_tokens": [ + -4.76391077041626 + ], + "logprob_reward": -0.013910956680774689, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013910956680774689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2383486032485962, + 0.09086757898330688, + 0.04501690715551376, + 0.03199443221092224, + 0.09871053695678711, + 0.0413442924618721, + 0.06084177643060684, + 0.04956435784697533, + 0.2546779215335846, + 0.08863364905118942 + ], + "credit_map": [ + 0.017409909516572952, + 0.028271790593862534, + 0.05118193477392197, + 0.015510338358581066, + 0.023836186155676842, + 0.020763471722602844, + 0.025836441665887833, + 0.05553851276636124, + 0.0075325313955545425, + 0.006233277264982462, + 0.027006879448890686, + 0.005211128853261471, + 0.006228665821254253, + 0.006073337979614735, + 0.008547447621822357, + 0.024034295231103897, + 0.004360042046755552, + 0.007205007132142782, + 0.007929167710244656, + 0.006716536357998848, + 0.0037118385080248117, + 0.004498632624745369, + 0.004236715380102396, + 0.006358962506055832, + 0.004278375301510096, + 0.004293723963201046, + 0.004964232910424471, + 0.003037047805264592, + 0.002470818581059575, + 0.0047671739012002945, + 0.003524778876453638, + 0.00465827900916338, + 0.010009550489485264, + 0.018240634351968765, + 0.014043537899851799, + 0.008924683555960655, + 0.013073031790554523, + 0.009036284871399403, + 0.012568934820592403, + 0.012813872657716274, + 0.0033544187899678946, + 0.007687455043196678, + 0.005282997619360685, + 0.0038944920524954796, + 0.0039259083569049835, + 0.007143212016671896, + 0.004247770644724369, + 0.005808032583445311, + 0.008122425526380539, + 0.007596784736961126, + 0.010070720687508583, + 0.00642685079947114, + 0.005342894699424505, + 0.00804063118994236, + 0.007830594666302204, + 0.0074108680710196495, + 0.0038295937702059746, + 0.0072265807539224625, + 0.013948873616755009, + 0.0037787496112287045, + 0.0037708887830376625, + 0.005115459207445383, + 0.004115071147680283, + 0.007779129780828953, + 0.032539475709199905, + 0.022007223218679428, + 0.06668662279844284, + 0.022012516856193542, + 0.017555754631757736, + 0.03804153949022293, + 0.01952523924410343, + 0.03630954027175903, + 0.009689494967460632, + 0.008350166492164135, + 0.0155335683375597, + 0.01033489778637886, + 0.007662942633032799, + 0.008575561456382275, + 0.014996679499745369, + 0.013490322045981884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.013910956680774689 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063539505004883 + ], + "opposite_logprob_tokens": [ + -0.28135401010513306 + ], + "logprob_reward": -1.4063539505004883, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063539505004883, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09206365793943405, + 0.23818418383598328, + 0.09752064198255539, + 0.13888251781463623, + 0.057156361639499664, + 0.06735245138406754, + 0.08687808364629745, + 0.12610754370689392, + 0.028769895434379578, + 0.06708470731973648 + ], + "credit_map": [ + 0.009003477171063423, + 0.017940562218427658, + 0.011105102486908436, + 0.010403181426227093, + 0.011538594961166382, + 0.011690476909279823, + 0.009038664400577545, + 0.011343594640493393, + 0.014955105260014534, + 0.017738783732056618, + 0.08002296835184097, + 0.00977481622248888, + 0.016974017024040222, + 0.00900854729115963, + 0.013529523275792599, + 0.07618042826652527, + 0.007322690449655056, + 0.02161482535302639, + 0.01336317416280508, + 0.010250847786664963, + 0.014554127119481564, + 0.006373625248670578, + 0.01125763077288866, + 0.012783716432750225, + 0.019448762759566307, + 0.021618932485580444, + 0.014604589901864529, + 0.012691055424511433, + 0.018275707960128784, + 0.015186569653451443, + 0.016177842393517494, + 0.020879054442048073, + 0.007939410395920277, + 0.008816186338663101, + 0.00808972492814064, + 0.0048245773650705814, + 0.006066509056836367, + 0.007166366558521986, + 0.0066239978186786175, + 0.007629591040313244, + 0.007338596507906914, + 0.01019627507776022, + 0.006300716660916805, + 0.006447660271078348, + 0.006300515960901976, + 0.014117845334112644, + 0.006782547105103731, + 0.009868294931948185, + 0.008007877506315708, + 0.01604464463889599, + 0.01102249976247549, + 0.010172352194786072, + 0.008802038617432117, + 0.007964295335114002, + 0.015599371865391731, + 0.009264999069273472, + 0.0064806342124938965, + 0.007767154835164547, + 0.03591226041316986, + 0.00811004638671875, + 0.006596447434276342, + 0.03508998081088066, + 0.010757487267255783, + 0.015393533743917942, + 0.0037202283274382353, + 0.0032335601281374693, + 0.006135351024568081, + 0.002563463058322668, + 0.002735086716711521, + 0.0034889960661530495, + 0.0023586677853018045, + 0.004534540697932243, + 0.0074829100631177425, + 0.008347189985215664, + 0.010604629293084145, + 0.009665385819971561, + 0.006148016545921564, + 0.008608000352978706, + 0.006956079509109259, + 0.009272498078644276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -1.4063539505004883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023319169878959656 + ], + "opposite_logprob_tokens": [ + -3.7733192443847656 + ], + "logprob_reward": -0.023319169878959656, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023319169878959656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10131224244832993, + 0.5473273396492004, + 0.03711332380771637, + 0.06975774466991425, + 0.05593889579176903, + 0.02907443791627884, + 0.024012181907892227, + 0.030169842764735222, + 0.04650238901376724, + 0.05879152938723564 + ], + "credit_map": [ + 0.012104799039661884, + 0.008047033101320267, + 0.015520418994128704, + 0.005282326601445675, + 0.007949440740048885, + 0.025850439444184303, + 0.011179091408848763, + 0.015378699637949467, + 0.03236817196011543, + 0.016952568665146828, + 0.207886204123497, + 0.022763527929782867, + 0.01782291941344738, + 0.024271242320537567, + 0.02417621947824955, + 0.20108647644519806, + 0.004295118153095245, + 0.005282985512167215, + 0.00786447711288929, + 0.003439381718635559, + 0.003525857347995043, + 0.003829345339909196, + 0.0035341691691428423, + 0.005341989919543266, + 0.010565565899014473, + 0.009388749487698078, + 0.00955918151885271, + 0.005525501444935799, + 0.007248071487993002, + 0.011336613446474075, + 0.005446697119623423, + 0.010687370784580708, + 0.004925766959786415, + 0.010607166215777397, + 0.009227604605257511, + 0.004936699755489826, + 0.005198921076953411, + 0.006435023155063391, + 0.007103241514414549, + 0.007504476699978113, + 0.0031341207213699818, + 0.0028609123546630144, + 0.003721503308042884, + 0.0026258830912411213, + 0.0019970496650785208, + 0.0068653603084385395, + 0.002919829683378339, + 0.004949777852743864, + 0.0026488678995519876, + 0.0034548367839306593, + 0.0038918633945286274, + 0.002834941027686, + 0.002251388505101204, + 0.0028995375614613295, + 0.003571851411834359, + 0.002458894858136773, + 0.0017610280774533749, + 0.0015533843543380499, + 0.011551405303180218, + 0.0017922479892149568, + 0.0015969306696206331, + 0.006746586412191391, + 0.0024393510539084673, + 0.0027289085555821657, + 0.005251098889857531, + 0.004545377567410469, + 0.0071982331573963165, + 0.004353759810328484, + 0.0032602851279079914, + 0.011079540476202965, + 0.0039937435649335384, + 0.00682035181671381, + 0.006948547903448343, + 0.006856427062302828, + 0.009018772281706333, + 0.008116478100419044, + 0.004713727626949549, + 0.01001062709838152, + 0.005093847401440144, + 0.00803310051560402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.023319169878959656 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648592788726091 + ], + "opposite_logprob_tokens": [ + -5.379648685455322 + ], + "logprob_reward": -0.004648592788726091, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648592788726091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15221089124679565, + 0.2787962853908539, + 0.07662301510572433, + 0.05424542725086212, + 0.10223879665136337, + 0.04447588324546814, + 0.03574659675359726, + 0.07715760916471481, + 0.11352824419736862, + 0.0649772584438324 + ], + "credit_map": [ + 0.018323751166462898, + 0.018193010240793228, + 0.02604627050459385, + 0.01495321374386549, + 0.015486563555896282, + 0.02540448307991028, + 0.013407420367002487, + 0.020396174862980843, + 0.017552131786942482, + 0.027169689536094666, + 0.08874291926622391, + 0.012572109699249268, + 0.017808301374316216, + 0.01188666932284832, + 0.017989477142691612, + 0.08507497608661652, + 0.0077941990457475185, + 0.01265727262943983, + 0.01281349454075098, + 0.008760650642216206, + 0.008216671645641327, + 0.008012275211513042, + 0.007082780823111534, + 0.011285672895610332, + 0.008468805812299252, + 0.008599231019616127, + 0.005433559417724609, + 0.0051469276659190655, + 0.00610694894567132, + 0.006270527373999357, + 0.005890152417123318, + 0.008329271338880062, + 0.008285612799227238, + 0.016701247543096542, + 0.020108412951231003, + 0.010421848855912685, + 0.011049906723201275, + 0.007035109214484692, + 0.01315345149487257, + 0.015483200550079346, + 0.003997135441750288, + 0.0065375021658837795, + 0.007031883113086224, + 0.0047712624073028564, + 0.003635026980191469, + 0.007418872322887182, + 0.004729722626507282, + 0.00635447958484292, + 0.00444813771173358, + 0.006588398478925228, + 0.005262188613414764, + 0.0031141084618866444, + 0.0030424478463828564, + 0.0056235953234136105, + 0.004715770483016968, + 0.0029519505333155394, + 0.004241173155605793, + 0.004245261196047068, + 0.028896022588014603, + 0.0043809679336845875, + 0.004096058662980795, + 0.01756558194756508, + 0.007933573797345161, + 0.005798973608762026, + 0.012180113233625889, + 0.011545158922672272, + 0.027181100100278854, + 0.012255177833139896, + 0.010794682428240776, + 0.014412965625524521, + 0.010063796304166317, + 0.015095241367816925, + 0.008014895021915436, + 0.006287809927016497, + 0.0073816594667732716, + 0.008611252531409264, + 0.005250736605376005, + 0.011820741929113865, + 0.007905516773462296, + 0.009704641997814178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.004648592788726091 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779046289622784 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009779046289622784, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779046289622784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12476944178342819, + 0.19917145371437073, + 0.056310854852199554, + 0.06269257515668869, + 0.08515693992376328, + 0.031764548271894455, + 0.04062044620513916, + 0.11223553866147995, + 0.19968350231647491, + 0.08759471029043198 + ], + "credit_map": [ + 0.009565533138811588, + 0.019687367603182793, + 0.023645034059882164, + 0.011500667780637741, + 0.013149549253284931, + 0.015931932255625725, + 0.015206471085548401, + 0.0160828810185194, + 0.013887791894376278, + 0.015511516481637955, + 0.06870588660240173, + 0.00912391021847725, + 0.00797660555690527, + 0.009841826744377613, + 0.01020105741918087, + 0.06392286717891693, + 0.004659266211092472, + 0.011400947347283363, + 0.009528289549052715, + 0.005780317820608616, + 0.0054074982181191444, + 0.00513030169531703, + 0.007285145577043295, + 0.007119093090295792, + 0.00711795361712575, + 0.00971787329763174, + 0.00686514750123024, + 0.005843639373779297, + 0.007561524398624897, + 0.00803222693502903, + 0.007120547816157341, + 0.01043365802615881, + 0.007579123135656118, + 0.013751148246228695, + 0.012798149138689041, + 0.007231677416712046, + 0.00917673297226429, + 0.011801626533269882, + 0.010722996667027473, + 0.012095482088625431, + 0.0033356929197907448, + 0.004117855802178383, + 0.004856565035879612, + 0.003330129198729992, + 0.002702870173379779, + 0.004755144473165274, + 0.0037670466117560863, + 0.004899246152490377, + 0.004227419849485159, + 0.004863150883466005, + 0.004876001738011837, + 0.005123981740325689, + 0.003792467759922147, + 0.006309599615633488, + 0.006084604654461145, + 0.005343218334019184, + 0.00654566241428256, + 0.006679563783109188, + 0.04073707014322281, + 0.006962708197534084, + 0.007519461680203676, + 0.022889554500579834, + 0.012390115298330784, + 0.008511406369507313, + 0.025408433750271797, + 0.027120348066091537, + 0.02708088420331478, + 0.02391701564192772, + 0.019912797957658768, + 0.03278808295726776, + 0.015948019921779633, + 0.027507923543453217, + 0.007972138933837414, + 0.013988319784402847, + 0.012263249605894089, + 0.009123759344220161, + 0.007696241606026888, + 0.0077831498347222805, + 0.016702143475413322, + 0.012065708637237549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.009779046289622784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, the camera held still, captures a bird foraging on the quiet, moonlit forest floor. Cool moonlight filters through the sparse branches of tall trees, casting dappled shadows on a carpet of dry, curled leaves. The bird's plumage, covered in a complex array of brown, black, and beige feathers, makes it blend almost seamlessly into its surroundings. Its large, round, black eyes gleam with alertness in the dim light, while its extremely long, slender beak is intensely focused on its task. With steady, precise care, it plunges its beak into the pile of dead leaves at its feet, then rhythmically probes up and down, each dip accompanied by the subtle rustling of leaves being gently brushed aside. The moonlight outlines the contours of its back feathers and a silver highlight on its beak. In the background, several moss-covered tree roots meander in deep shadows. The whole scene, with the main colors of dark blue and earthy brown, presents an extremely peaceful, focused and slightly mysterious atmosphere.", + "seed": 5321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_0/index_5671_seed_5321_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2216623889282346, + "vlm_logprob": -0.2216623889282346, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.2216623889282346, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03033391386270523 + ], + "opposite_logprob_tokens": [ + -3.905333995819092 + ], + "logprob_reward": -0.03033391386270523, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03033391386270523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20517361164093018, + 0.05780773237347603, + 0.04001208394765854, + 0.2079552859067917, + 0.05101434513926506, + 0.12102881073951721, + 0.07006631791591644, + 0.10831855237483978, + 0.0697137638926506, + 0.0689094290137291 + ], + "credit_map": [ + 0.01797487586736679, + 0.007302341517060995, + 0.04733380302786827, + 0.016182612627744675, + 0.014502211473882198, + 0.028171204030513763, + 0.012638555839657784, + 0.06106803938746452, + 0.0039608231745660305, + 0.0030735821928828955, + 0.006537051405757666, + 0.005757550243288279, + 0.009539175778627396, + 0.005695682484656572, + 0.015170406550168991, + 0.008073462173342705, + 0.0032276147976517677, + 0.004241021350026131, + 0.005154243670403957, + 0.003751261392608285, + 0.006536843720823526, + 0.0059624649584293365, + 0.005958918947726488, + 0.0051797181367874146, + 0.011126542463898659, + 0.013694140128791332, + 0.02879091165959835, + 0.025114037096500397, + 0.03558214008808136, + 0.027229469269514084, + 0.033997032791376114, + 0.03242102637887001, + 0.0036810715682804585, + 0.0029575543012470007, + 0.008381666615605354, + 0.008480831049382687, + 0.007874511182308197, + 0.005701033398509026, + 0.006964872591197491, + 0.006972808390855789, + 0.006430322770029306, + 0.006322578061372042, + 0.019160764291882515, + 0.018472857773303986, + 0.019272781908512115, + 0.010515783913433552, + 0.02040669322013855, + 0.02044704370200634, + 0.004635836463421583, + 0.0045319125056266785, + 0.010505640879273415, + 0.006990666501224041, + 0.013051587156951427, + 0.01141719426959753, + 0.008668104186654091, + 0.010265388526022434, + 0.007494301535189152, + 0.004575169645249844, + 0.015087159350514412, + 0.013622353784739971, + 0.019688762724399567, + 0.01582181639969349, + 0.01733807660639286, + 0.014690925367176533, + 0.004539427347481251, + 0.004456170368939638, + 0.011979250237345695, + 0.012120531871914864, + 0.00943179614841938, + 0.008655907586216927, + 0.007543181534856558, + 0.010987501591444016, + 0.004261924885213375, + 0.0047700353898108006, + 0.012339171953499317, + 0.009364578872919083, + 0.01157646719366312, + 0.006837504915893078, + 0.009882381185889244, + 0.009877369739115238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.03033391386270523 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539332628250122 + ], + "opposite_logprob_tokens": [ + -0.4289332628250122 + ], + "logprob_reward": -1.0539332628250122, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539332628250122, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10375303030014038, + 0.06866828352212906, + 0.04419146850705147, + 0.14782221615314484, + 0.15193071961402893, + 0.13556987047195435, + 0.07527363300323486, + 0.0911940485239029, + 0.10832314193248749, + 0.07327350229024887 + ], + "credit_map": [ + 0.011239727027714252, + 0.005383661948144436, + 0.01874929666519165, + 0.015424475073814392, + 0.010872328653931618, + 0.015245764516294003, + 0.010706858709454536, + 0.01613093540072441, + 0.005295703187584877, + 0.0036241640336811543, + 0.008121313527226448, + 0.007112654857337475, + 0.008906523697078228, + 0.009637503884732723, + 0.016533980146050453, + 0.009436450898647308, + 0.003871419932693243, + 0.004953006282448769, + 0.004780828021466732, + 0.00549346162006259, + 0.009643599390983582, + 0.00463043712079525, + 0.005243863444775343, + 0.0055748638696968555, + 0.00789621565490961, + 0.00540895015001297, + 0.019151218235492706, + 0.015697157010436058, + 0.036290243268013, + 0.01713588647544384, + 0.026176804676651955, + 0.020065777003765106, + 0.007103857584297657, + 0.0059657590463757515, + 0.02420992963016033, + 0.020823929458856583, + 0.02705259621143341, + 0.01050354354083538, + 0.03300704061985016, + 0.023264095187187195, + 0.00909226480871439, + 0.00857651885598898, + 0.020139776170253754, + 0.021842358633875847, + 0.026042388752102852, + 0.013024610467255116, + 0.01831311173737049, + 0.018538866192102432, + 0.006504072807729244, + 0.004565196577459574, + 0.00995324831455946, + 0.006164622958749533, + 0.014028768055140972, + 0.015210510231554508, + 0.00753408670425415, + 0.011313138529658318, + 0.007212989963591099, + 0.003830286441370845, + 0.012120014056563377, + 0.014474490657448769, + 0.0167018361389637, + 0.012868613004684448, + 0.01149874459952116, + 0.012487092055380344, + 0.008482467383146286, + 0.004830831661820412, + 0.015858106315135956, + 0.021563062444329262, + 0.02038675732910633, + 0.012341059744358063, + 0.010991404764354229, + 0.013869470916688442, + 0.007228170521557331, + 0.005109517835080624, + 0.010907338932156563, + 0.008417320437729359, + 0.011679337359964848, + 0.012168308719992638, + 0.00840919278562069, + 0.00935432594269514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -1.0539332628250122 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008657186292111874 + ], + "opposite_logprob_tokens": [ + -4.758656978607178 + ], + "logprob_reward": -0.008657186292111874, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008657186292111874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35862138867378235, + 0.056472182273864746, + 0.05381651595234871, + 0.08973904699087143, + 0.054065145552158356, + 0.055789440870285034, + 0.10229731351137161, + 0.11643671244382858, + 0.04633840173482895, + 0.06642387807369232 + ], + "credit_map": [ + 0.03502542898058891, + 0.018213491886854172, + 0.07610592246055603, + 0.038759320974349976, + 0.02619132213294506, + 0.04738246276974678, + 0.02005244418978691, + 0.09689100831747055, + 0.008983638137578964, + 0.004657073877751827, + 0.006968430709093809, + 0.006697497796267271, + 0.005314266309142113, + 0.009935806505382061, + 0.006223093718290329, + 0.007692372892051935, + 0.004294093698263168, + 0.002545658964663744, + 0.00844914373010397, + 0.005067432764917612, + 0.011513523757457733, + 0.008757039904594421, + 0.006030082702636719, + 0.007159537170082331, + 0.005721485707908869, + 0.004788535181432962, + 0.013034427538514137, + 0.009768933057785034, + 0.02194126322865486, + 0.007132801692932844, + 0.014241661876440048, + 0.013109940104186535, + 0.004172447137534618, + 0.003159071085974574, + 0.010456090793013573, + 0.007384696509689093, + 0.008020486682653427, + 0.006759575102478266, + 0.006676872260868549, + 0.00743590435013175, + 0.00712974090129137, + 0.0036387373693287373, + 0.00860337819904089, + 0.0065702409483492374, + 0.009198044426739216, + 0.0064095910638570786, + 0.006734000518918037, + 0.007505708374083042, + 0.008297901600599289, + 0.00897460337728262, + 0.020107656717300415, + 0.010638297535479069, + 0.018163226544857025, + 0.008860192261636257, + 0.011784411035478115, + 0.015471025370061398, + 0.013081020675599575, + 0.006571851205080748, + 0.020914603024721146, + 0.012964513152837753, + 0.014764540828764439, + 0.016675101593136787, + 0.012653660029172897, + 0.01881142146885395, + 0.0023736257571727037, + 0.002484223572537303, + 0.006983160972595215, + 0.008450972847640514, + 0.010257750749588013, + 0.005119766108691692, + 0.004259609151631594, + 0.006409292574971914, + 0.004671423230320215, + 0.0034997262991964817, + 0.011005493812263012, + 0.005270054098218679, + 0.009490115568041801, + 0.013050672598183155, + 0.007226034067571163, + 0.012210354208946228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.008657186292111874 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865293201059103 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.00865293201059103, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865293201059103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1753399521112442, + 0.05663067102432251, + 0.07196658849716187, + 0.16347390413284302, + 0.0972207635641098, + 0.07778458297252655, + 0.06678006052970886, + 0.0973503366112709, + 0.09585894644260406, + 0.09759417176246643 + ], + "credit_map": [ + 0.016695188358426094, + 0.009663031436502934, + 0.03110682964324951, + 0.018351813778281212, + 0.025908423587679863, + 0.02072902023792267, + 0.014574370346963406, + 0.03831126540899277, + 0.0050633312202990055, + 0.003698489861562848, + 0.0070244004018604755, + 0.006938887760043144, + 0.00906505435705185, + 0.009089911356568336, + 0.008578031323850155, + 0.0071725668385624886, + 0.004734689835458994, + 0.004691290203481913, + 0.009356686845421791, + 0.009918810799717903, + 0.013845687732100487, + 0.00997956469655037, + 0.00954390037804842, + 0.00989595428109169, + 0.010458527132868767, + 0.010078535415232182, + 0.02370833419263363, + 0.029729068279266357, + 0.027977006509900093, + 0.012540440075099468, + 0.026176709681749344, + 0.022805267944931984, + 0.00530413119122386, + 0.004239467438310385, + 0.015148043632507324, + 0.012728850357234478, + 0.022426825016736984, + 0.009032213129103184, + 0.014605971984565258, + 0.013735258020460606, + 0.006020842120051384, + 0.0033190546091645956, + 0.01179193053394556, + 0.00989123061299324, + 0.018858375027775764, + 0.007261107675731182, + 0.009708071127533913, + 0.01093396544456482, + 0.004636695608496666, + 0.004211423918604851, + 0.009927800856530666, + 0.007026462350040674, + 0.012123270891606808, + 0.00944824330508709, + 0.009798943996429443, + 0.009607214480638504, + 0.005292685702443123, + 0.00436052493751049, + 0.016061890870332718, + 0.014541801996529102, + 0.020042670890688896, + 0.009440168738365173, + 0.012865347787737846, + 0.014745243825018406, + 0.006223883479833603, + 0.004602195229381323, + 0.01405943650752306, + 0.019416743889451027, + 0.015358787961304188, + 0.010526622645556927, + 0.012640413828194141, + 0.013030868954956532, + 0.00985575933009386, + 0.006128002889454365, + 0.013515838421881199, + 0.011214718222618103, + 0.01554617565125227, + 0.018358906731009483, + 0.009044116362929344, + 0.013930657878518105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.00865293201059103 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734649650752544 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734649650752544, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734649650752544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25134965777397156, + 0.0814339742064476, + 0.03473983332514763, + 0.0710846409201622, + 0.11539067327976227, + 0.04035806283354759, + 0.1036563292145729, + 0.09960901737213135, + 0.15056940913200378, + 0.051808346062898636 + ], + "credit_map": [ + 0.020645592361688614, + 0.011420153081417084, + 0.061937347054481506, + 0.020398693159222603, + 0.015886757522821426, + 0.030167849734425545, + 0.01568654179573059, + 0.07520673424005508, + 0.011161244474351406, + 0.006440665107220411, + 0.008759495802223682, + 0.009194490499794483, + 0.007390011567622423, + 0.016743473708629608, + 0.010419061407446861, + 0.011325537227094173, + 0.0031417161226272583, + 0.002305349800735712, + 0.005007013212889433, + 0.003738886909559369, + 0.005258888937532902, + 0.0036276564933359623, + 0.006700129248201847, + 0.004960195627063513, + 0.005503649357706308, + 0.0035112714394927025, + 0.013554025441408157, + 0.009235575795173645, + 0.007851923815906048, + 0.008994407020509243, + 0.012564145028591156, + 0.009869654662907124, + 0.005218741483986378, + 0.004554274491965771, + 0.018165817484259605, + 0.01841888576745987, + 0.02211480401456356, + 0.009622897952795029, + 0.022779393941164017, + 0.014515874907374382, + 0.0030140611343085766, + 0.0020279812160879374, + 0.005392753519117832, + 0.0052202423103153706, + 0.008127343840897083, + 0.00520178722217679, + 0.005228092428296804, + 0.006145806983113289, + 0.006974701303988695, + 0.00483533414080739, + 0.01860259845852852, + 0.008972637355327606, + 0.014300120994448662, + 0.010469133034348488, + 0.022553695365786552, + 0.016948115080595016, + 0.007772184442728758, + 0.0037939734756946564, + 0.014521358534693718, + 0.011320617981255054, + 0.017973419278860092, + 0.01356604415923357, + 0.01912512257695198, + 0.011536302976310253, + 0.0062814573757350445, + 0.0061057135462760925, + 0.0201187115162611, + 0.029145950451493263, + 0.03714277967810631, + 0.013951724395155907, + 0.015724334865808487, + 0.022098759189248085, + 0.004136793781071901, + 0.003410380333662033, + 0.006570233963429928, + 0.007235495373606682, + 0.008647557348012924, + 0.008014710620045662, + 0.007192741148173809, + 0.0066004348918795586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.006734649650752544 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a low angle, the camera held still, captures a bird foraging on the quiet, moonlit forest floor. Cool moonlight filters through the sparse branches of tall trees, casting dappled shadows on a carpet of dry, curled leaves. The bird's plumage, covered in a complex array of brown, black, and beige feathers, makes it blend almost seamlessly into its surroundings. Its large, round, black eyes gleam with alertness in the dim light, while its extremely long, slender beak is intensely focused on its task. With steady, precise care, it plunges its beak into the pile of dead leaves at its feet, then rhythmically probes up and down, each dip accompanied by the subtle rustling of leaves being gently brushed aside. The moonlight outlines the contours of its back feathers and a silver highlight on its beak. In the background, several moss-covered tree roots meander in deep shadows. The whole scene, with the main colors of dark blue and earthy brown, presents an extremely peaceful, focused and slightly mysterious atmosphere.", + "seed": 5322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_0/index_5671_seed_5322_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1730315494351089, + "vlm_logprob": -0.1730315494351089, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.1730315494351089, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01871054619550705 + ], + "opposite_logprob_tokens": [ + -4.518710613250732 + ], + "logprob_reward": -0.01871054619550705, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01871054619550705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12921084463596344, + 0.12062942236661911, + 0.07621465623378754, + 0.04995029792189598, + 0.17033323645591736, + 0.06804938614368439, + 0.1005847305059433, + 0.030290206894278526, + 0.1702362298965454, + 0.08450096845626831 + ], + "credit_map": [ + 0.01210739091038704, + 0.013428588397800922, + 0.021998515352606773, + 0.015306135639548302, + 0.014987417496740818, + 0.02101629227399826, + 0.014992850832641125, + 0.015373654663562775, + 0.009104481898248196, + 0.009559054858982563, + 0.024712201207876205, + 0.014019867405295372, + 0.014812015928328037, + 0.021007398143410683, + 0.019517691805958748, + 0.007896706461906433, + 0.0050660683773458, + 0.004097247961908579, + 0.010905788280069828, + 0.008087590336799622, + 0.017461339011788368, + 0.01322535052895546, + 0.0076096877455711365, + 0.00976158119738102, + 0.004313028883188963, + 0.0045646303333342075, + 0.008552063256502151, + 0.005744501948356628, + 0.008388133719563484, + 0.00743457255885005, + 0.006046058610081673, + 0.0049073053523898125, + 0.018381088972091675, + 0.01478398498147726, + 0.02502857707440853, + 0.015137244015932083, + 0.017915811389684677, + 0.028678441420197487, + 0.029869817197322845, + 0.02053827978670597, + 0.006858242675662041, + 0.006072985474020243, + 0.00945704523473978, + 0.007813587784767151, + 0.0083669014275074, + 0.012281540781259537, + 0.0066492315381765366, + 0.010549850761890411, + 0.00990967731922865, + 0.007713077124208212, + 0.01428481936454773, + 0.011884192936122417, + 0.00996016338467598, + 0.02154950611293316, + 0.011687853373587132, + 0.013595439493656158, + 0.003510404611006379, + 0.002856011036783457, + 0.0033728007692843676, + 0.003575459122657776, + 0.0040718745440244675, + 0.005796872079372406, + 0.004324111621826887, + 0.002782674739137292, + 0.011695227585732937, + 0.013858561404049397, + 0.020976513624191284, + 0.01943947747349739, + 0.030579067766666412, + 0.02488747611641884, + 0.027631990611553192, + 0.02116791531443596, + 0.006468910723924637, + 0.0069206710904836655, + 0.010705317370593548, + 0.011754130944609642, + 0.014932097867131233, + 0.013066324405372143, + 0.010061843320727348, + 0.010591672733426094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.01871054619550705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.826221227645874 + ], + "opposite_logprob_tokens": [ + -0.576221227645874 + ], + "logprob_reward": -0.826221227645874, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.826221227645874, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20280222594738007, + 0.24414026737213135, + 0.05034608393907547, + 0.055529315024614334, + 0.04360703378915787, + 0.066268190741539, + 0.08727951347827911, + 0.05223161354660988, + 0.158725306391716, + 0.03907041251659393 + ], + "credit_map": [ + 0.014162692241370678, + 0.014698939397931099, + 0.031420670449733734, + 0.029797255992889404, + 0.029000617563724518, + 0.031097305938601494, + 0.02827547676861286, + 0.02434930019080639, + 0.021594326943159103, + 0.018516309559345245, + 0.03876425698399544, + 0.026973163709044456, + 0.03413805365562439, + 0.04100044444203377, + 0.02867223508656025, + 0.03448150306940079, + 0.004063984379172325, + 0.0036832340992987156, + 0.006464231293648481, + 0.00384602346457541, + 0.011759321205317974, + 0.00807257927954197, + 0.008342495188117027, + 0.004114223178476095, + 0.0056928060948848724, + 0.005609214771538973, + 0.007634441833943129, + 0.007682694122195244, + 0.010373013094067574, + 0.007310349494218826, + 0.0070872013457119465, + 0.004139602184295654, + 0.0031803972087800503, + 0.002829755889251828, + 0.005187581293284893, + 0.006918162573128939, + 0.005777406506240368, + 0.007582411635667086, + 0.007774977944791317, + 0.004356349818408489, + 0.005485242232680321, + 0.003969703335314989, + 0.007780951447784901, + 0.011168223805725574, + 0.013293839059770107, + 0.008786645717918873, + 0.010411682538688183, + 0.005371910519897938, + 0.00626890454441309, + 0.00538275483995676, + 0.01094644796103239, + 0.00848383642733097, + 0.010524974204599857, + 0.020653191953897476, + 0.009655235335230827, + 0.01536418404430151, + 0.006172982510179281, + 0.006525083910673857, + 0.0058237770572304726, + 0.0062585510313510895, + 0.004809929057955742, + 0.007514073047786951, + 0.007588211912661791, + 0.007539017125964165, + 0.013833939097821712, + 0.013365218415856361, + 0.02129322476685047, + 0.021485794335603714, + 0.016694732010364532, + 0.019860444590449333, + 0.01649749092757702, + 0.03569449856877327, + 0.0025186222046613693, + 0.002370670437812805, + 0.004840297624468803, + 0.005877770483493805, + 0.006397304590791464, + 0.007229247130453587, + 0.005684602074325085, + 0.0041519044898450375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.826221227645874 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746490485966206 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746490485966206, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746490485966206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16006770730018616, + 0.24988272786140442, + 0.11659114062786102, + 0.043110109865665436, + 0.11369573324918747, + 0.0863659679889679, + 0.06295110285282135, + 0.043629854917526245, + 0.08516820520162582, + 0.038537461310625076 + ], + "credit_map": [ + 0.014197595417499542, + 0.015352836810052395, + 0.025043802335858345, + 0.020284373313188553, + 0.015963193029165268, + 0.025265974923968315, + 0.019894184544682503, + 0.024065731093287468, + 0.02049419842660427, + 0.020184127613902092, + 0.038786645978689194, + 0.03166314214468002, + 0.03645000606775284, + 0.04138048738241196, + 0.02572586201131344, + 0.03519822657108307, + 0.012237032875418663, + 0.014029053039848804, + 0.016434049233794212, + 0.011969808489084244, + 0.016854962334036827, + 0.018872249871492386, + 0.011516190133988857, + 0.014677785336971283, + 0.004209467209875584, + 0.0060011777095496655, + 0.005091032013297081, + 0.005940002389252186, + 0.007236298639327288, + 0.004809418227523565, + 0.0053441934287548065, + 0.0044785174541175365, + 0.01930970326066017, + 0.020821254700422287, + 0.0161041971296072, + 0.008356181904673576, + 0.009842334315180779, + 0.01298527792096138, + 0.013165634125471115, + 0.013111134991049767, + 0.004952843300998211, + 0.005168160889297724, + 0.01045099925249815, + 0.009891347959637642, + 0.014031171798706055, + 0.015781205147504807, + 0.009375303983688354, + 0.016714928671717644, + 0.006004178896546364, + 0.00625995546579361, + 0.007245280314236879, + 0.010448387823998928, + 0.008705445565283298, + 0.008881282992661, + 0.007265076506882906, + 0.008141489699482918, + 0.004837984684854746, + 0.005065858364105225, + 0.005245127249509096, + 0.005905970931053162, + 0.0044867354445159435, + 0.006360959727317095, + 0.005979578942060471, + 0.0057476358488202095, + 0.008594452403485775, + 0.006801946088671684, + 0.00949416309595108, + 0.013049282133579254, + 0.012222524732351303, + 0.011284509673714638, + 0.012817347422242165, + 0.010903974063694477, + 0.002963672624900937, + 0.003775242483243346, + 0.004553194157779217, + 0.004670911002904177, + 0.007667974103242159, + 0.007210908457636833, + 0.004005667753517628, + 0.0036898853722959757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.006746490485966206 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745188031345606 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745188031345606, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745188031345606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15630501508712769, + 0.2385481894016266, + 0.0880635529756546, + 0.03396757319569588, + 0.09010037034749985, + 0.04421418160200119, + 0.06532935053110123, + 0.033651962876319885, + 0.18472537398338318, + 0.06509443372488022 + ], + "credit_map": [ + 0.013292953372001648, + 0.013745540753006935, + 0.022189384326338768, + 0.017521880567073822, + 0.021441277116537094, + 0.027169248089194298, + 0.021912427619099617, + 0.019032292068004608, + 0.0181560218334198, + 0.01727725937962532, + 0.03738022968173027, + 0.02620580978691578, + 0.04044121876358986, + 0.04093840718269348, + 0.037724751979112625, + 0.02042449079453945, + 0.006420518737286329, + 0.0060509489849209785, + 0.023301824927330017, + 0.006223777309060097, + 0.011357123963534832, + 0.021720822900533676, + 0.008061639033257961, + 0.004926903173327446, + 0.002842607442289591, + 0.002688483102247119, + 0.0051407660357654095, + 0.003826362546533346, + 0.0043857283890247345, + 0.00567985512316227, + 0.005475517828017473, + 0.0039282552897930145, + 0.010922574438154697, + 0.011411231011152267, + 0.012901692651212215, + 0.008328222669661045, + 0.007906115613877773, + 0.013371603563427925, + 0.015507135540246964, + 0.009751793928444386, + 0.0036227444652467966, + 0.0033677418250590563, + 0.005698658060282469, + 0.005320013500750065, + 0.00581367639824748, + 0.008869051933288574, + 0.0057496000081300735, + 0.005772698670625687, + 0.007731451187282801, + 0.0060027725994586945, + 0.008335578255355358, + 0.006799944210797548, + 0.006004228722304106, + 0.01210696715861559, + 0.009664367884397507, + 0.00868404284119606, + 0.005583090241998434, + 0.004262780770659447, + 0.003798076882958412, + 0.004345469176769257, + 0.003134296042844653, + 0.003226961474865675, + 0.003964380826801062, + 0.005336909554898739, + 0.0176792461425066, + 0.01902679167687893, + 0.022055119276046753, + 0.030614301562309265, + 0.023743990808725357, + 0.021427059546113014, + 0.015727154910564423, + 0.03445171192288399, + 0.0038317288272082806, + 0.004367690533399582, + 0.008036825805902481, + 0.008139540441334248, + 0.013033969327807426, + 0.012698374688625336, + 0.0075162616558372974, + 0.007470038253813982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.006745188031345606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a bird foraging on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit by cool moonlight filtering through tall trees with dappled shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird have brown, black, and beige feathers with a long slender beak and large round black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the bird's action shown as plunging and rhythmically probing its beak into a pile of dry, curled leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734294816851616 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734294816851616, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734294816851616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15252681076526642, + 0.10597167164087296, + 0.12758731842041016, + 0.06540430337190628, + 0.08524424582719803, + 0.0977904349565506, + 0.08395829796791077, + 0.03505922853946686, + 0.10523539781570435, + 0.1412222981452942 + ], + "credit_map": [ + 0.007666442543268204, + 0.01227967906743288, + 0.0189963448792696, + 0.020593302324414253, + 0.01874598301947117, + 0.029022550210356712, + 0.016084233298897743, + 0.02913827635347843, + 0.008794091641902924, + 0.00798893254250288, + 0.016226129606366158, + 0.012735623866319656, + 0.017527764663100243, + 0.017848877236247063, + 0.014007196761667728, + 0.010843055322766304, + 0.008189747110009193, + 0.008987885899841785, + 0.03624868020415306, + 0.010919361375272274, + 0.013138746842741966, + 0.032472796738147736, + 0.00972136203199625, + 0.00790874008089304, + 0.003834554459899664, + 0.0046494342386722565, + 0.012050755321979523, + 0.009221373125910759, + 0.008509576320648193, + 0.010446666739881039, + 0.009841425344347954, + 0.006850518751889467, + 0.00972940493375063, + 0.011693526990711689, + 0.013012695126235485, + 0.006240177433937788, + 0.00898119155317545, + 0.012709257192909718, + 0.014238174073398113, + 0.008639814332127571, + 0.007596861571073532, + 0.007082267664372921, + 0.013130230829119682, + 0.012321322225034237, + 0.0180822666734457, + 0.01877407357096672, + 0.011400424875319004, + 0.009402980096638203, + 0.010376796126365662, + 0.00963059812784195, + 0.010050785727798939, + 0.007786675356328487, + 0.009489941410720348, + 0.014295880682766438, + 0.009727947413921356, + 0.012599673122167587, + 0.0031901835463941097, + 0.0027578803710639477, + 0.003820073325186968, + 0.004829748999327421, + 0.005431140772998333, + 0.005805764347314835, + 0.005544146057218313, + 0.003680290188640356, + 0.009581591002643108, + 0.010171111673116684, + 0.012925132177770138, + 0.016503000631928444, + 0.016088007017970085, + 0.012670977972447872, + 0.011888284236192703, + 0.015407302416861057, + 0.009457777254283428, + 0.010467814281582832, + 0.01904839463531971, + 0.017963681370019913, + 0.025032266974449158, + 0.022509563714265823, + 0.019683143123984337, + 0.017059652134776115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.006734294816851616 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_1/index_tensor([5538, 5538])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_1/index_tensor([5538, 5538])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..355c460474765e07156c0b7c0fbae50bed7a3817 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_1/index_tensor([5538, 5538])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "drone shot, shot vertically from above, descends slowly and steadily in a calm, documentary style, capturing a single object floating on the vast, deep blue expanse of the open ocean. It's a weathered, square wooden box, its dark wood damp and cracked from prolonged immersion in seawater. Its edges are worn smooth, held together by a few rusty metal bands, even adorned with scattered green algae. It rises and falls rhythmically with the gentle waves, propelled steadily and slowly by an invisible current, leaving behind a faint, almost inconspicuous wake. Soft, diffuse sunlight penetrates the thin cloud layer, evenly scattering across the ocean surface. The box's damp surface reflects a faint luster, contrasting with the deep, almost inky waters surrounding it. The tiny ripples on the surface resemble scattered diamonds. The entire picture presents an extreme sense of tranquility and loneliness, with cold tones, emphasizing the vastness of the ocean and the smallness of this mysterious drifting object, creating an unresolved atmosphere full of story imagination.", + "seed": 1546, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_1/index_5538_seed_1546_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13330259090289473, + "vlm_logprob": -0.13330259090289473, + "vlm_yes_no_margin": 3.824999976158142, + "vlm_token_credit": -0.13330259090289473, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16814717650413513 + ], + "opposite_logprob_tokens": [ + -2.168147087097168 + ], + "logprob_reward": -0.16814717650413513, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.16814717650413513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15869292616844177, + 0.0672510489821434, + 0.049601443111896515, + 0.07165374606847763, + 0.07376756519079208, + 0.2665690779685974, + 0.10981201380491257, + 0.07038304954767227, + 0.08675755560398102, + 0.04551158845424652 + ], + "credit_map": [ + 0.018434075638651848, + 0.021656008437275887, + 0.02184605598449707, + 0.014555156230926514, + 0.020305605605244637, + 0.026840515434741974, + 0.015065459534525871, + 0.019990071654319763, + 0.012228835374116898, + 0.010549486614763737, + 0.008879936300218105, + 0.006016085855662823, + 0.006995710078626871, + 0.012129576876759529, + 0.00492666894569993, + 0.005524749401956797, + 0.00811854749917984, + 0.005838193465024233, + 0.007423289120197296, + 0.006115967873483896, + 0.004145797807723284, + 0.0076447464525699615, + 0.0048564220778644085, + 0.00545848160982132, + 0.01704544387757778, + 0.011045190505683422, + 0.012564612552523613, + 0.006783126853406429, + 0.004625246860086918, + 0.008733267895877361, + 0.0037152462173253298, + 0.0071416147984564304, + 0.0063573638908565044, + 0.012241565622389317, + 0.014492872171103954, + 0.00717023154720664, + 0.00553332082927227, + 0.009986583143472672, + 0.006262599024921656, + 0.011723027564585209, + 0.02587704360485077, + 0.028582200407981873, + 0.032735761255025864, + 0.03298389911651611, + 0.02097287029027939, + 0.040856506675481796, + 0.03505123034119606, + 0.04950956627726555, + 0.019272198900580406, + 0.015399807132780552, + 0.01875067502260208, + 0.011354217305779457, + 0.007766892202198505, + 0.013957327231764793, + 0.013341100886464119, + 0.009969801642000675, + 0.013860059902071953, + 0.012999708764255047, + 0.01056006457656622, + 0.009345745667815208, + 0.005064380820840597, + 0.006045014597475529, + 0.005657250992953777, + 0.006850820500403643, + 0.010423093102872372, + 0.008904565125703812, + 0.017271561548113823, + 0.019415877759456635, + 0.007073254324495792, + 0.0054452926851809025, + 0.009264267981052399, + 0.00895964726805687, + 0.00828615203499794, + 0.0028751399368047714, + 0.009989715181291103, + 0.004401960875838995, + 0.0029732300899922848, + 0.004579838830977678, + 0.005525592714548111, + 0.006879962049424648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.16814717650413513 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743252992630005 + ], + "opposite_logprob_tokens": [ + -0.9743252992630005 + ], + "logprob_reward": -0.4743252992630005, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743252992630005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17285461723804474, + 0.04424059018492699, + 0.056857455521821976, + 0.09168379753828049, + 0.1175810769200325, + 0.060802608728408813, + 0.030617104843258858, + 0.1496085524559021, + 0.2118544727563858, + 0.06389966607093811 + ], + "credit_map": [ + 0.03255807235836983, + 0.03541730344295502, + 0.020200733095407486, + 0.011329947970807552, + 0.0127970976755023, + 0.0415462926030159, + 0.00852112751454115, + 0.010484057478606701, + 0.004380617290735245, + 0.007154420483857393, + 0.007384427357465029, + 0.00505518214777112, + 0.003961336333304644, + 0.00821797363460064, + 0.0036266259849071503, + 0.0044600083492696285, + 0.007695085369050503, + 0.0070032598450779915, + 0.011163006536662579, + 0.006377584300935268, + 0.004717062693089247, + 0.008931335993111134, + 0.0047837416641414165, + 0.006186380051076412, + 0.02099827490746975, + 0.016259653493762016, + 0.011174637824296951, + 0.010042000561952591, + 0.006125131156295538, + 0.009662389755249023, + 0.006597814150154591, + 0.01082390546798706, + 0.016682356595993042, + 0.021925311535596848, + 0.019759224727749825, + 0.012364523485302925, + 0.008286315947771072, + 0.0130329430103302, + 0.010794932022690773, + 0.01473547238856554, + 0.00840243510901928, + 0.006471469532698393, + 0.007475752383470535, + 0.00858409982174635, + 0.004274855367839336, + 0.007882056757807732, + 0.005848668050020933, + 0.01186327077448368, + 0.004684734623879194, + 0.0037572390865534544, + 0.0036728358827531338, + 0.002536113141104579, + 0.0035936932545155287, + 0.004380808211863041, + 0.004746173042804003, + 0.0032455073669552803, + 0.023157142102718353, + 0.01805661804974079, + 0.020358461886644363, + 0.015348311513662338, + 0.015978164970874786, + 0.014941141940653324, + 0.02345316857099533, + 0.01831553876399994, + 0.032684870064258575, + 0.021397734060883522, + 0.04343494772911072, + 0.05379263311624527, + 0.010329362004995346, + 0.010552817024290562, + 0.01901758834719658, + 0.020644523203372955, + 0.012189707718789577, + 0.004556338768452406, + 0.015358084812760353, + 0.008721188642084599, + 0.0045472197234630585, + 0.004882494453340769, + 0.006332069635391235, + 0.007312560919672251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.4743252992630005 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015195977175608277 + ], + "opposite_logprob_tokens": [ + -6.501519680023193 + ], + "logprob_reward": -0.0015195977175608277, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015195977175608277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15196099877357483, + 0.07092273235321045, + 0.04155554622411728, + 0.12809212505817413, + 0.06030964106321335, + 0.08399268239736557, + 0.09808307886123657, + 0.13305385410785675, + 0.02744004875421524, + 0.20458939671516418 + ], + "credit_map": [ + 0.021848801523447037, + 0.024877559393644333, + 0.017021553590893745, + 0.012260244227945805, + 0.013745015487074852, + 0.033503226935863495, + 0.015060342848300934, + 0.013644225895404816, + 0.00673292251303792, + 0.008251109160482883, + 0.009014500305056572, + 0.00986532960087061, + 0.006703196093440056, + 0.010903851129114628, + 0.010776706039905548, + 0.008675108663737774, + 0.0031370571814477444, + 0.00425766222178936, + 0.007157693617045879, + 0.0060026878491044044, + 0.004229369573295116, + 0.006151236593723297, + 0.003990933299064636, + 0.006628898438066244, + 0.011995881795883179, + 0.016502518206834793, + 0.020823834463953972, + 0.014683816581964493, + 0.009876233525574207, + 0.02721361629664898, + 0.012189806438982487, + 0.014806391671299934, + 0.008128339424729347, + 0.007309899199754, + 0.011194514110684395, + 0.008928587660193443, + 0.005684456322342157, + 0.005830088164657354, + 0.0061198873445391655, + 0.007113857660442591, + 0.007391707971692085, + 0.009425156749784946, + 0.010288561694324017, + 0.012531526386737823, + 0.0056501468643546104, + 0.01159659493714571, + 0.010471079498529434, + 0.0166378915309906, + 0.015843192115426064, + 0.015590742230415344, + 0.014469609595835209, + 0.01105191744863987, + 0.008274074643850327, + 0.010840386152267456, + 0.012476451694965363, + 0.009536686353385448, + 0.013870053924620152, + 0.010586998425424099, + 0.017874902114272118, + 0.016165517270565033, + 0.00966811552643776, + 0.012704274617135525, + 0.02056913636624813, + 0.03161483258008957, + 0.003781213192269206, + 0.003268711967393756, + 0.00420427368953824, + 0.003262631129473448, + 0.003000149969011545, + 0.0031558305490761995, + 0.0034253853373229504, + 0.0033418487291783094, + 0.0332053005695343, + 0.0128283416852355, + 0.04775853082537651, + 0.026685193181037903, + 0.012740680016577244, + 0.02150518260896206, + 0.02239091508090496, + 0.027475208044052124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0015195977175608277 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001943010720424354 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001943010720424354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001943010720424354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13008396327495575, + 0.08763672411441803, + 0.07377266138792038, + 0.08227875828742981, + 0.0806812271475792, + 0.09015777707099915, + 0.07939205318689346, + 0.07974709570407867, + 0.0413484200835228, + 0.2549014091491699 + ], + "credit_map": [ + 0.01972157508134842, + 0.016700321808457375, + 0.01403970830142498, + 0.011540207080543041, + 0.012524188496172428, + 0.02572716400027275, + 0.015244284644722939, + 0.014586501754820347, + 0.006829370278865099, + 0.009201198816299438, + 0.011726140975952148, + 0.011059623211622238, + 0.00947942677885294, + 0.017389947548508644, + 0.011713417246937752, + 0.010237587615847588, + 0.009563433937728405, + 0.010679326951503754, + 0.011091900989413261, + 0.008075462654232979, + 0.006735799368470907, + 0.01225867960602045, + 0.006920936983078718, + 0.008447108790278435, + 0.0076208533719182014, + 0.010856898501515388, + 0.014160875231027603, + 0.010026802308857441, + 0.006038835272192955, + 0.016467006877064705, + 0.007386514917016029, + 0.009720969013869762, + 0.009313936345279217, + 0.010405538603663445, + 0.012561236508190632, + 0.011652546934783459, + 0.007216690108180046, + 0.008386055938899517, + 0.007154884748160839, + 0.013990326784551144, + 0.013837534934282303, + 0.010891882702708244, + 0.015043728053569794, + 0.011433116160333157, + 0.004935836419463158, + 0.011423235759139061, + 0.007929517887532711, + 0.01466291956603527, + 0.013819627463817596, + 0.012953895144164562, + 0.012928972952067852, + 0.008388803340494633, + 0.006597048137336969, + 0.008742637000977993, + 0.00802244246006012, + 0.007938623428344727, + 0.007165177725255489, + 0.007103120442479849, + 0.009170611388981342, + 0.009033441543579102, + 0.0072351680137217045, + 0.009414647705852985, + 0.013421911746263504, + 0.01720300503075123, + 0.005769617389887571, + 0.0039944774471223354, + 0.00787111185491085, + 0.0088829156011343, + 0.003302475903183222, + 0.003317248309031129, + 0.004805037751793861, + 0.003405533032491803, + 0.04159891605377197, + 0.01756124757230282, + 0.05507900193333626, + 0.030273644253611565, + 0.021034393459558487, + 0.02766793966293335, + 0.025905484333634377, + 0.03578074276447296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.001943010720424354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577870309352875 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577870309352875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577870309352875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14647702872753143, + 0.03981034830212593, + 0.08496298640966415, + 0.17756125330924988, + 0.05342641845345497, + 0.10042700171470642, + 0.0886131301522255, + 0.05060616508126259, + 0.15124158561229706, + 0.10687407851219177 + ], + "credit_map": [ + 0.030881471931934357, + 0.023932816460728645, + 0.01879689283668995, + 0.0072804102674126625, + 0.014514259994029999, + 0.030148539692163467, + 0.010418872348964214, + 0.010503767989575863, + 0.005745782516896725, + 0.005781130865216255, + 0.0061438786797225475, + 0.004683057311922312, + 0.003914548549801111, + 0.005227416753768921, + 0.004319818690419197, + 0.003994707949459553, + 0.01151243131607771, + 0.010816036723554134, + 0.011590448208153248, + 0.008829054422676563, + 0.007464843336492777, + 0.013606812804937363, + 0.008321321569383144, + 0.012822030112147331, + 0.046911340206861496, + 0.036326561123132706, + 0.018016338348388672, + 0.011599586345255375, + 0.014972630888223648, + 0.023554859682917595, + 0.01280019711703062, + 0.013379731215536594, + 0.008315159939229488, + 0.009355584159493446, + 0.006894701160490513, + 0.004773653578013182, + 0.005477759521454573, + 0.00553432572633028, + 0.0052164834924042225, + 0.00785874854773283, + 0.008534202352166176, + 0.01097009889781475, + 0.012509401887655258, + 0.01368215773254633, + 0.00781838409602642, + 0.015228953212499619, + 0.012838541530072689, + 0.01884525641798973, + 0.022970804944634438, + 0.01299054641276598, + 0.02063220553100109, + 0.007436021231114864, + 0.006711302325129509, + 0.006825102027505636, + 0.005472867749631405, + 0.005574280396103859, + 0.009341012686491013, + 0.005854502785950899, + 0.006230696104466915, + 0.005135567393153906, + 0.006336540449410677, + 0.005720584187656641, + 0.006869438569992781, + 0.005117821041494608, + 0.02341601997613907, + 0.01351260393857956, + 0.026137152686715126, + 0.03138120099902153, + 0.01081653218716383, + 0.009423822164535522, + 0.021548761054873466, + 0.015005480498075485, + 0.02652961201965809, + 0.00986980926245451, + 0.028885511681437492, + 0.009472664445638657, + 0.005788495298475027, + 0.006811385974287987, + 0.009606250561773777, + 0.00991034135222435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.020577870309352875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone shot, shot vertically from above, descends slowly and steadily in a calm, documentary style, capturing a single object floating on the vast, deep blue expanse of the open ocean. It's a weathered, square wooden box, its dark wood damp and cracked from prolonged immersion in seawater. Its edges are worn smooth, held together by a few rusty metal bands, even adorned with scattered green algae. It rises and falls rhythmically with the gentle waves, propelled steadily and slowly by an invisible current, leaving behind a faint, almost inconspicuous wake. Soft, diffuse sunlight penetrates the thin cloud layer, evenly scattering across the ocean surface. The box's damp surface reflects a faint luster, contrasting with the deep, almost inky waters surrounding it. The tiny ripples on the surface resemble scattered diamonds. The entire picture presents an extreme sense of tranquility and loneliness, with cold tones, emphasizing the vastness of the ocean and the smallness of this mysterious drifting object, creating an unresolved atmosphere full of story imagination.", + "seed": 1547, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_1/index_5538_seed_1547_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15831622290425001, + "vlm_logprob": -0.15831622290425001, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.15831622290425001, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29080405831336975 + ], + "opposite_logprob_tokens": [ + -1.5408040285110474 + ], + "logprob_reward": -0.29080405831336975, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.29080405831336975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14877985417842865, + 0.08039316534996033, + 0.03451061248779297, + 0.0480390302836895, + 0.48252904415130615, + 0.026567818596959114, + 0.029941953718662262, + 0.012111609801650047, + 0.11963112652301788, + 0.01749570667743683 + ], + "credit_map": [ + 0.015931783244013786, + 0.01207895390689373, + 0.017273688688874245, + 0.022483890876173973, + 0.014687158167362213, + 0.026904037222266197, + 0.01987309381365776, + 0.01954725570976734, + 0.007356785237789154, + 0.005667909048497677, + 0.014731908217072487, + 0.012814264744520187, + 0.006173314992338419, + 0.01096316333860159, + 0.011929984204471111, + 0.010755845345556736, + 0.0032333582639694214, + 0.0033510313369333744, + 0.0052663483656942844, + 0.00546821067109704, + 0.003092467552050948, + 0.004352062940597534, + 0.005048953462392092, + 0.004698181990534067, + 0.005233719013631344, + 0.004477576352655888, + 0.006929803639650345, + 0.006761775352060795, + 0.0041364505887031555, + 0.005448171403259039, + 0.0051483833231031895, + 0.009903151541948318, + 0.03610171005129814, + 0.015833871439099312, + 0.1735411137342453, + 0.014698212966322899, + 0.01866178773343563, + 0.11122116446495056, + 0.007910619489848614, + 0.10456060618162155, + 0.0024562578182667494, + 0.0020535036455839872, + 0.00472229951992631, + 0.002730973996222019, + 0.002166129881516099, + 0.0033750825095921755, + 0.00248127244412899, + 0.006582302041351795, + 0.0026473356410861015, + 0.002637495519593358, + 0.005767488852143288, + 0.004337359219789505, + 0.002860204316675663, + 0.002437770599499345, + 0.0040159462951123714, + 0.005238355603069067, + 0.0016398398438468575, + 0.0012090964009985328, + 0.0016456296434625983, + 0.001112273894250393, + 0.0012732280883938074, + 0.002068827161565423, + 0.0010122674284502864, + 0.00215044803917408, + 0.02600996010005474, + 0.009327181614935398, + 0.012713088653981686, + 0.01059529185295105, + 0.017270488664507866, + 0.013356444425880909, + 0.010332169011235237, + 0.020026518031954765, + 0.002011731034144759, + 0.0019199955277144909, + 0.0023093554191291332, + 0.002514014020562172, + 0.0026489535812288523, + 0.0015373592032119632, + 0.002227535704150796, + 0.0023267657961696386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.29080405831336975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743259847164154 + ], + "opposite_logprob_tokens": [ + -0.9743260145187378 + ], + "logprob_reward": -0.4743259847164154, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743259847164154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04899986460804939, + 0.029089275747537613, + 0.032470520585775375, + 0.04398900270462036, + 0.6200805902481079, + 0.015395968221127987, + 0.05823877081274986, + 0.013702547177672386, + 0.028729211539030075, + 0.10930425673723221 + ], + "credit_map": [ + 0.005854669492691755, + 0.006058298982679844, + 0.0072657219134271145, + 0.007368452847003937, + 0.004527102690190077, + 0.005730610806494951, + 0.006744849029928446, + 0.0054501560516655445, + 0.003627847647294402, + 0.0025691278278827667, + 0.005856802221387625, + 0.0023837455082684755, + 0.002709530061110854, + 0.004819680005311966, + 0.002138864016160369, + 0.0049836765974760056, + 0.0036056225653737783, + 0.002864984329789877, + 0.0050096516497433186, + 0.004517572466284037, + 0.0022617564536631107, + 0.005626147147268057, + 0.004184843506664038, + 0.004399938974529505, + 0.006491417530924082, + 0.0028206631541252136, + 0.005034243687987328, + 0.004011827986687422, + 0.0027629942633211613, + 0.010579450987279415, + 0.0038554633501917124, + 0.00843294057995081, + 0.048813022673130035, + 0.020461732521653175, + 0.21802550554275513, + 0.018573220819234848, + 0.025037815794348717, + 0.1525227278470993, + 0.009703573770821095, + 0.12694299221038818, + 0.001789806061424315, + 0.0013301268918439746, + 0.0020223951432853937, + 0.0022821747697889805, + 0.0013662000419571996, + 0.001949445460923016, + 0.001621707109734416, + 0.003034112276509404, + 0.004432033747434616, + 0.0038817934691905975, + 0.012353884056210518, + 0.007541441824287176, + 0.004108838737010956, + 0.004255932755768299, + 0.007793311029672623, + 0.013871531002223492, + 0.0014862845418974757, + 0.0014663280453532934, + 0.0016772777307778597, + 0.001840810407884419, + 0.0014705610228702426, + 0.0018814589129760861, + 0.0015371419722214341, + 0.0023426832631230354, + 0.003529591718688607, + 0.003179267980158329, + 0.002903544809669256, + 0.0034285790752619505, + 0.0046408758498728275, + 0.004654019605368376, + 0.002259014407172799, + 0.0041343155317008495, + 0.010073467157781124, + 0.011329014785587788, + 0.013207138516008854, + 0.01449466310441494, + 0.021452581509947777, + 0.009982346557080746, + 0.01617482677102089, + 0.012590205296874046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.4743259847164154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645982291549444 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.004645982291549444, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645982291549444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05251647159457207, + 0.023232970386743546, + 0.020387331023812294, + 0.022004036232829094, + 0.7516965866088867, + 0.014686649665236473, + 0.0650797113776207, + 0.0028112875297665596, + 0.026958588510751724, + 0.02062634937465191 + ], + "credit_map": [ + 0.005657218396663666, + 0.005728956311941147, + 0.006188490893691778, + 0.010504379868507385, + 0.004148685373365879, + 0.008362013846635818, + 0.00601497245952487, + 0.005911755375564098, + 0.002263010712340474, + 0.0030336438212543726, + 0.0036053636576980352, + 0.00383113999851048, + 0.002034686040133238, + 0.0029156042728573084, + 0.0023220416624099016, + 0.0032274790573865175, + 0.0019297059625387192, + 0.002117769094184041, + 0.0027682518120855093, + 0.0031583954114466906, + 0.0015960866585373878, + 0.0030321571975946426, + 0.0030499566346406937, + 0.002735006157308817, + 0.002353721996769309, + 0.0012178578181192279, + 0.00285789230838418, + 0.001465371809899807, + 0.0014378639170899987, + 0.005715565290302038, + 0.0014245890779420733, + 0.005531171802431345, + 0.05925299599766731, + 0.02485833503305912, + 0.26369228959083557, + 0.021685486659407616, + 0.030407197773456573, + 0.18372975289821625, + 0.013434710912406445, + 0.15463578701019287, + 0.0014744673389941454, + 0.0016380796441808343, + 0.0017947430023923516, + 0.002376048592850566, + 0.0015991555992513895, + 0.0017245030030608177, + 0.0019545448012650013, + 0.002125105122104287, + 0.0028044141363352537, + 0.0014373089652508497, + 0.025528602302074432, + 0.0022930181585252285, + 0.001274770824238658, + 0.0030318063218146563, + 0.001551418681629002, + 0.02715836651623249, + 0.0002675015421118587, + 0.00031851750100031495, + 0.0004312456294428557, + 0.0004523807729128748, + 0.00029116173391230404, + 0.00033229877590201795, + 0.00031462317565456033, + 0.00040355819510295987, + 0.005426560528576374, + 0.0022651462350040674, + 0.0028472421690821648, + 0.0027622277848422527, + 0.0037763817235827446, + 0.003125489456579089, + 0.002591341268271208, + 0.0041641960851848125, + 0.0020757613237947226, + 0.002363630337640643, + 0.0023121279664337635, + 0.0023901225067675114, + 0.004805729258805513, + 0.002103158039972186, + 0.0022127535194158554, + 0.002363064559176564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.004645982291549444 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003618955845013261 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.003618955845013261, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003618955845013261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046522676944732666, + 0.02004547417163849, + 0.01787414401769638, + 0.018907133489847183, + 0.8062009811401367, + 0.015394114889204502, + 0.03289913386106491, + 0.006010628305375576, + 0.023464752361178398, + 0.012681011110544205 + ], + "credit_map": [ + 0.0049354881048202515, + 0.004885988309979439, + 0.007882599718868732, + 0.00753006199374795, + 0.0036163614131510258, + 0.005643101409077644, + 0.006684258580207825, + 0.005344810429960489, + 0.0020127426832914352, + 0.0034248116426169872, + 0.0026786935050040483, + 0.003478723345324397, + 0.001699342392385006, + 0.002186034806072712, + 0.002525488380342722, + 0.002039635321125388, + 0.0018695410108193755, + 0.002118214499205351, + 0.0024665258824825287, + 0.0029496888164430857, + 0.0015087744686752558, + 0.002404712373390794, + 0.0021832867059856653, + 0.002373398281633854, + 0.0016510240966454148, + 0.0010799841256812215, + 0.0025202352553606033, + 0.0013144157128408551, + 0.0013389608357101679, + 0.0045353686437010765, + 0.001345486263744533, + 0.0051216562278568745, + 0.06421410292387009, + 0.02672341652214527, + 0.28246167302131653, + 0.02296304516494274, + 0.03275054693222046, + 0.20017153024673462, + 0.013762970454990864, + 0.16315366327762604, + 0.0014862136449664831, + 0.0013377993600443006, + 0.001973746344447136, + 0.00228540669195354, + 0.0014254961861297488, + 0.0019298355327919126, + 0.0025374344550073147, + 0.002418181160464883, + 0.0015098861185833812, + 0.0006194491870701313, + 0.01297709345817566, + 0.0010302270529791713, + 0.0007081073126755655, + 0.0015842856373637915, + 0.0007336334674619138, + 0.01373644731938839, + 0.0006265685078687966, + 0.0005723267095163465, + 0.0009117184090428054, + 0.0007417694432660937, + 0.0006813079817220569, + 0.0007787481299601495, + 0.0008136950782500207, + 0.000884493172634393, + 0.003689250210300088, + 0.0021428854670375586, + 0.002406488638371229, + 0.0020234647672623396, + 0.003401449415832758, + 0.004907404538244009, + 0.0015984323108568788, + 0.003295374335721135, + 0.001246389583684504, + 0.001369125791825354, + 0.001958767417818308, + 0.0021634739823639393, + 0.0020230975933372974, + 0.0011195200495421886, + 0.0015391181223094463, + 0.001261517172679305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.003618955845013261 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186133354902267 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018186133354902267, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186133354902267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.009327742271125317, + 0.007806377485394478, + 0.008583617396652699, + 0.03352240473031998, + 0.9103690385818481, + 0.011450349353253841, + 0.008092538453638554, + 0.0022838925942778587, + 0.004718448501080275, + 0.0038456544280052185 + ], + "credit_map": [ + 0.00109842827077955, + 0.0010545434197410941, + 0.001504784799180925, + 0.0010151044698432088, + 0.0009627485414966941, + 0.0010787282371893525, + 0.0014848406426608562, + 0.001128563773818314, + 0.0007104381802491844, + 0.0010533558670431376, + 0.0012782863341271877, + 0.0011171485530212522, + 0.0006142197526060045, + 0.0009783883579075336, + 0.0011527742026373744, + 0.0009017666452564299, + 0.0006739635718986392, + 0.000880409381352365, + 0.0013219095999374986, + 0.001165480469353497, + 0.0009185843518935144, + 0.0011881970567628741, + 0.0014084808062762022, + 0.001026591518893838, + 0.0025441439356654882, + 0.0030475787352770567, + 0.004644918721169233, + 0.003905274672433734, + 0.0026437987107783556, + 0.005458711180835962, + 0.0030283615924417973, + 0.008249618113040924, + 0.07098184525966644, + 0.03002435527741909, + 0.3223961591720581, + 0.025994788855314255, + 0.03618280962109566, + 0.22075577080249786, + 0.01407739706337452, + 0.1899559199810028, + 0.0010377181461080909, + 0.0009595483425073326, + 0.0017159548588097095, + 0.0013791965320706367, + 0.0012026942567899823, + 0.0012447199551388621, + 0.00147340539842844, + 0.002437111921608448, + 0.00045028486056253314, + 0.0003615661698859185, + 0.0028421091847121716, + 0.00044057611376047134, + 0.00025976335746236145, + 0.00043318202369846404, + 0.0004122815153095871, + 0.002892774762585759, + 0.00024308564024977386, + 0.0002243383787572384, + 0.0003256879572290927, + 0.0003029480285476893, + 0.00020962281269021332, + 0.0003477443242445588, + 0.0002553681842982769, + 0.0003750970063265413, + 0.0010412713745608926, + 0.0004450563283171505, + 0.0004965708358213305, + 0.00048586944467388093, + 0.0005847053253091872, + 0.0005165012553334236, + 0.0004431512497831136, + 0.0007053225417621434, + 0.0003991317644249648, + 0.0004393437993712723, + 0.000474877335363999, + 0.00044339237501844764, + 0.0006641997606493533, + 0.00043345571612007916, + 0.0004305410839151591, + 0.0005607125349342823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.018186133354902267 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone shot, shot vertically from above, descends slowly and steadily in a calm, documentary style, capturing a single object floating on the vast, deep blue expanse of the open ocean. It's a weathered, square wooden box, its dark wood damp and cracked from prolonged immersion in seawater. Its edges are worn smooth, held together by a few rusty metal bands, even adorned with scattered green algae. It rises and falls rhythmically with the gentle waves, propelled steadily and slowly by an invisible current, leaving behind a faint, almost inconspicuous wake. Soft, diffuse sunlight penetrates the thin cloud layer, evenly scattering across the ocean surface. The box's damp surface reflects a faint luster, contrasting with the deep, almost inky waters surrounding it. The tiny ripples on the surface resemble scattered diamonds. The entire picture presents an extreme sense of tranquility and loneliness, with cold tones, emphasizing the vastness of the ocean and the smallness of this mysterious drifting object, creating an unresolved atmosphere full of story imagination.", + "seed": 1548, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_1/index_5538_seed_1548_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11748413851018995, + "vlm_logprob": -0.11748413851018995, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.11748413851018995, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3370300531387329 + ], + "opposite_logprob_tokens": [ + -1.337030053138733 + ], + "logprob_reward": -0.3370300531387329, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3370300531387329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055975496768951416, + 0.01846267096698284, + 0.02677883580327034, + 0.010539880022406578, + 0.04790744557976723, + 0.012577669695019722, + 0.02528008446097374, + 0.03179086744785309, + 0.5274513959884644, + 0.24323566257953644 + ], + "credit_map": [ + 0.006999508943408728, + 0.006527738180011511, + 0.008770680986344814, + 0.006434460170567036, + 0.0060281772166490555, + 0.007441665977239609, + 0.005795822944492102, + 0.007977449335157871, + 0.0017860824009403586, + 0.00259093358181417, + 0.0028071156702935696, + 0.001982177374884486, + 0.001860947348177433, + 0.001986134797334671, + 0.0016486215172335505, + 0.00380066130310297, + 0.001888610771857202, + 0.002323329681530595, + 0.003396719926968217, + 0.0017613006057217717, + 0.0027092902455478907, + 0.006094463635236025, + 0.001720334286801517, + 0.00688478909432888, + 0.0013298357371240854, + 0.001112964702770114, + 0.0015920967562124133, + 0.0014717603335157037, + 0.0009427512995898724, + 0.0010671778582036495, + 0.0016609131125733256, + 0.0013623808044940233, + 0.004205921199172735, + 0.004330134484916925, + 0.00508904317393899, + 0.008852480910718441, + 0.004670315887778997, + 0.004258322529494762, + 0.009281087666749954, + 0.007220141123980284, + 0.0016163395484909415, + 0.0012819019611924887, + 0.0014310127589851618, + 0.0019210103200748563, + 0.0014581914292648435, + 0.001118055428378284, + 0.0022223221603780985, + 0.0015288356225937605, + 0.002929416950792074, + 0.0026616735849529505, + 0.0049584051594138145, + 0.0027750793378800154, + 0.002452022861689329, + 0.003231026930734515, + 0.0026246001943945885, + 0.0036478613037616014, + 0.0031736993696540594, + 0.003159502986818552, + 0.006579936947673559, + 0.0035872224252671003, + 0.0022481789346784353, + 0.0032045878469944, + 0.004368070513010025, + 0.005469668190926313, + 0.020308690145611763, + 0.008193294517695904, + 0.030180035158991814, + 0.02200574055314064, + 0.013062098063528538, + 0.15565168857574463, + 0.044627487659454346, + 0.23342235386371613, + 0.043442703783512115, + 0.019048994407057762, + 0.025091756135225296, + 0.026006575673818588, + 0.02713429182767868, + 0.03766368329524994, + 0.030280636623501778, + 0.034567058086395264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.3370300531387329 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560909390449524 + ], + "opposite_logprob_tokens": [ + -1.6006090641021729 + ], + "logprob_reward": -0.22560909390449524, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560909390449524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018293233588337898, + 0.012582087889313698, + 0.0278646070510149, + 0.007671153172850609, + 0.03460913896560669, + 0.008255049586296082, + 0.012129433453083038, + 0.016462214291095734, + 0.24366578459739685, + 0.6184673309326172 + ], + "credit_map": [ + 0.0025018895976245403, + 0.002681623212993145, + 0.00254086684435606, + 0.0019942596554756165, + 0.001786841661669314, + 0.0016645424766466022, + 0.0020887237042188644, + 0.0030344866681843996, + 0.0014927732991054654, + 0.0017322103958576918, + 0.0018131759716197848, + 0.0009976283181458712, + 0.0010700438870117068, + 0.0015234834281727672, + 0.000990897649899125, + 0.0029618756379932165, + 0.002185332588851452, + 0.0026079160161316395, + 0.0024575479328632355, + 0.001793030183762312, + 0.0018335141940042377, + 0.007928529754281044, + 0.0016688622999936342, + 0.007389875594526529, + 0.0009335022186860442, + 0.0008589558419771492, + 0.0011625341139733791, + 0.0008126092725433409, + 0.0008136674296110868, + 0.0008187760831788182, + 0.0011641737073659897, + 0.0011069347383454442, + 0.0027678594924509525, + 0.0030077267438173294, + 0.0038194628432393074, + 0.006983268540352583, + 0.0032236394472420216, + 0.0033091700170189142, + 0.00654821889474988, + 0.004949793219566345, + 0.0008227284415625036, + 0.0007508904673159122, + 0.0010106111876666546, + 0.0015264353714883327, + 0.0009211701108142734, + 0.0008273024577647448, + 0.0013003629865124822, + 0.001095548621378839, + 0.001444428344257176, + 0.0011282911291345954, + 0.002595050260424614, + 0.0015565671492367983, + 0.0011531207710504532, + 0.0013672829372808337, + 0.0012789021711796522, + 0.0016057903412729502, + 0.0020359805785119534, + 0.001328895683400333, + 0.002394289942458272, + 0.002496784320101142, + 0.0014327032258734107, + 0.002304789610207081, + 0.002282034605741501, + 0.002186736324802041, + 0.00878428015857935, + 0.0035082967951893806, + 0.013461541384458542, + 0.0105653190985322, + 0.005744333378970623, + 0.072012759745121, + 0.021396838128566742, + 0.10819242149591446, + 0.0874289721250534, + 0.04543018713593483, + 0.060960348695516586, + 0.06702481955289841, + 0.05905843898653984, + 0.09571351110935211, + 0.07561560720205307, + 0.12723542749881744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.22560909390449524 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002491467399522662 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.002491467399522662, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002491467399522662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16664960980415344, + 0.038587115705013275, + 0.058719784021377563, + 0.05906946212053299, + 0.0719832107424736, + 0.07123520225286484, + 0.019846485927700996, + 0.047098394483327866, + 0.3507286310195923, + 0.1160820797085762 + ], + "credit_map": [ + 0.013739013113081455, + 0.014558115042746067, + 0.037638258188962936, + 0.01964116282761097, + 0.010730442591011524, + 0.01345053780823946, + 0.014910932630300522, + 0.04198114946484566, + 0.0037159682251513004, + 0.005257334094494581, + 0.0063117630779743195, + 0.004602483008056879, + 0.004011848010122776, + 0.004858185537159443, + 0.003250389127060771, + 0.006579138804227114, + 0.004326985217630863, + 0.004431664943695068, + 0.005454777739942074, + 0.005005211569368839, + 0.003872587578371167, + 0.016637733206152916, + 0.0036244031507521868, + 0.015366420149803162, + 0.00790121778845787, + 0.013994975946843624, + 0.007660329807549715, + 0.005946969147771597, + 0.0062638879753649235, + 0.005621872376650572, + 0.005245186388492584, + 0.006435020826756954, + 0.007362363860011101, + 0.00665690703317523, + 0.009113054722547531, + 0.013619566336274147, + 0.006642297841608524, + 0.006735826376825571, + 0.009599030017852783, + 0.012254162691533566, + 0.006643260829150677, + 0.0071832891553640366, + 0.008779776282608509, + 0.01526413019746542, + 0.005724275950342417, + 0.006429852452129126, + 0.010944920592010021, + 0.010265694931149483, + 0.0022610644809901714, + 0.002582512330263853, + 0.0032667198684066534, + 0.0022197654470801353, + 0.001860244432464242, + 0.0019027510425075889, + 0.002496711676940322, + 0.0032567137386649847, + 0.005974546540528536, + 0.007014297880232334, + 0.004462814889848232, + 0.00783243402838707, + 0.005716611631214619, + 0.0051784515380859375, + 0.005195857025682926, + 0.005723377224057913, + 0.013457095250487328, + 0.004779116250574589, + 0.01981465518474579, + 0.01455067191272974, + 0.009337668307125568, + 0.1027696505188942, + 0.029991086572408676, + 0.1560286581516266, + 0.012904842384159565, + 0.010934360325336456, + 0.016050396487116814, + 0.019502103328704834, + 0.010909020900726318, + 0.0125407250598073, + 0.01487667579203844, + 0.0183639507740736 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.002491467399522662 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017144280718639493 + ], + "opposite_logprob_tokens": [ + -6.37671422958374 + ], + "logprob_reward": -0.0017144280718639493, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017144280718639493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2324942946434021, + 0.04942159354686737, + 0.04045213758945465, + 0.05226102098822594, + 0.07150956243276596, + 0.050517454743385315, + 0.016600612550973892, + 0.09218534827232361, + 0.12812864780426025, + 0.2664293050765991 + ], + "credit_map": [ + 0.01683749258518219, + 0.015985913574695587, + 0.06711208820343018, + 0.01957128569483757, + 0.009920977987349033, + 0.016734711825847626, + 0.012000887654721737, + 0.07433094084262848, + 0.004629252012819052, + 0.005789381451904774, + 0.008385669440031052, + 0.00460572587326169, + 0.004162175115197897, + 0.008143185637891293, + 0.0036973876412957907, + 0.010008816607296467, + 0.0038590379990637302, + 0.0039190854877233505, + 0.00485847657546401, + 0.0034048156812787056, + 0.003442354965955019, + 0.009497878141701221, + 0.0029845002572983503, + 0.008485985919833183, + 0.0070953890681266785, + 0.012169945053756237, + 0.006829288322478533, + 0.004899195861071348, + 0.005807309877127409, + 0.0051299287006258965, + 0.004495587665587664, + 0.0058343783020973206, + 0.0061967442743480206, + 0.006894963793456554, + 0.009389393962919712, + 0.012665406800806522, + 0.007655036635696888, + 0.006787565536797047, + 0.01157141663134098, + 0.010349038988351822, + 0.004335111007094383, + 0.004741589538753033, + 0.005968322511762381, + 0.011280610226094723, + 0.003991931676864624, + 0.004120281897485256, + 0.00881281215697527, + 0.007266795728355646, + 0.001861476106569171, + 0.0021621426567435265, + 0.0021765665151178837, + 0.002108888002112508, + 0.0019638072699308395, + 0.002156791277229786, + 0.002181853400543332, + 0.001989086391404271, + 0.01024684589356184, + 0.006030654069036245, + 0.01748920977115631, + 0.015132782980799675, + 0.007067093160003424, + 0.011755092069506645, + 0.011689037084579468, + 0.012774634175002575, + 0.00569506362080574, + 0.0030764818657189608, + 0.008234428241848946, + 0.005975042004138231, + 0.004548332653939724, + 0.03511468693614006, + 0.010722668841481209, + 0.05476194620132446, + 0.034028321504592896, + 0.01931104250252247, + 0.027603207156062126, + 0.03292929008603096, + 0.023505184799432755, + 0.04045812413096428, + 0.032543376088142395, + 0.05605076253414154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0017144280718639493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057565003633499 + ], + "opposite_logprob_tokens": [ + -3.895575761795044 + ], + "logprob_reward": -0.02057565003633499, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057565003633499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12782087922096252, + 0.020341694355010986, + 0.019330738112330437, + 0.015005484223365784, + 0.02696072869002819, + 0.019423913210630417, + 0.012083357200026512, + 0.01298372820019722, + 0.4728931486606598, + 0.27315637469291687 + ], + "credit_map": [ + 0.008738900534808636, + 0.007053332403302193, + 0.03567071259021759, + 0.0065641216933727264, + 0.00558551074936986, + 0.014338156208395958, + 0.007388858124613762, + 0.04248128458857536, + 0.0022597394417971373, + 0.003109781537204981, + 0.002600999316200614, + 0.0018123140325769782, + 0.002881032880395651, + 0.003424222581088543, + 0.0014174475800246, + 0.0028361575677990913, + 0.0016674615908414125, + 0.0015772728947922587, + 0.0024785404093563557, + 0.001310311141423881, + 0.0013367197243496776, + 0.004733970854431391, + 0.001394079765304923, + 0.0048323823139071465, + 0.0017163987504318357, + 0.001725453999824822, + 0.0019848321098834276, + 0.0019434745190665126, + 0.0017517250962555408, + 0.001439768704585731, + 0.002492655534297228, + 0.0019511753926053643, + 0.0033999024890363216, + 0.002021097345277667, + 0.0039573232643306255, + 0.004925233777612448, + 0.0024496230762451887, + 0.002948642475530505, + 0.003910288214683533, + 0.003348617348819971, + 0.0016566556878387928, + 0.0020929044112563133, + 0.002605143468827009, + 0.003198759164661169, + 0.0015122003387659788, + 0.002202120143920183, + 0.00322520243935287, + 0.0029309268575161695, + 0.0011911082547158003, + 0.001259678858332336, + 0.0019128104904666543, + 0.0016468033427372575, + 0.0014869313454255462, + 0.0013777241110801697, + 0.0016722697764635086, + 0.001536031486466527, + 0.0016863427590578794, + 0.0012579511385411024, + 0.0020131426863372326, + 0.001704938244074583, + 0.0013289741473272443, + 0.0016644005663692951, + 0.0017333407886326313, + 0.0015946381026878953, + 0.01919965073466301, + 0.012320533394813538, + 0.030138710513710976, + 0.024265367537736893, + 0.014305665157735348, + 0.13374817371368408, + 0.03816620260477066, + 0.20074884593486786, + 0.03659061715006828, + 0.01861671544611454, + 0.02704976685345173, + 0.030601246282458305, + 0.02464972622692585, + 0.042370788753032684, + 0.03389069437980652, + 0.05938682332634926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.02057565003633499 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "drone shot, shot vertically from above, descends slowly and steadily in a calm, documentary style, capturing a single object floating on the vast, deep blue expanse of the open ocean. It's a weathered, square wooden box, its dark wood damp and cracked from prolonged immersion in seawater. Its edges are worn smooth, held together by a few rusty metal bands, even adorned with scattered green algae. It rises and falls rhythmically with the gentle waves, propelled steadily and slowly by an invisible current, leaving behind a faint, almost inconspicuous wake. Soft, diffuse sunlight penetrates the thin cloud layer, evenly scattering across the ocean surface. The box's damp surface reflects a faint luster, contrasting with the deep, almost inky waters surrounding it. The tiny ripples on the surface resemble scattered diamonds. The entire picture presents an extreme sense of tranquility and loneliness, with cold tones, emphasizing the vastness of the ocean and the smallness of this mysterious drifting object, creating an unresolved atmosphere full of story imagination.", + "seed": 1549, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_1/index_5538_seed_1549_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08678861437365412, + "vlm_logprob": -0.08678861437365412, + "vlm_yes_no_margin": 3.749999976158142, + "vlm_token_credit": -0.08678861437365412, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18369494378566742 + ], + "opposite_logprob_tokens": [ + -2.058694839477539 + ], + "logprob_reward": -0.18369494378566742, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.18369494378566742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11455267667770386, + 0.07647241652011871, + 0.04680034890770912, + 0.06457652151584625, + 0.02763860486447811, + 0.07114488631486893, + 0.23713061213493347, + 0.2591237425804138, + 0.0231038685888052, + 0.07945629954338074 + ], + "credit_map": [ + 0.015325724147260189, + 0.010190067812800407, + 0.01742817834019661, + 0.009420742280781269, + 0.014937106519937515, + 0.018476491793990135, + 0.011562822386622429, + 0.0172115508466959, + 0.005891052540391684, + 0.009080252610147, + 0.01119243074208498, + 0.013065100647509098, + 0.008882379159331322, + 0.009416508488357067, + 0.009309014305472374, + 0.009635685943067074, + 0.0053113894537091255, + 0.004560975823551416, + 0.006388809531927109, + 0.00545985484495759, + 0.006787083111703396, + 0.005066499579697847, + 0.007558322045952082, + 0.005667415913194418, + 0.009418367408216, + 0.005057984963059425, + 0.00893996562808752, + 0.0090522775426507, + 0.0056778062134981155, + 0.011715609580278397, + 0.006090066395699978, + 0.008624443784356117, + 0.0039330231957137585, + 0.0028330530039966106, + 0.0037726324517279863, + 0.00321212294511497, + 0.0035241490695625544, + 0.004690269939601421, + 0.0023723402991890907, + 0.003301014192402363, + 0.007615850772708654, + 0.005759172607213259, + 0.009241664782166481, + 0.007776971440762281, + 0.005626322701573372, + 0.012083517387509346, + 0.01259750034660101, + 0.010443886741995811, + 0.018870722502470016, + 0.025572773069143295, + 0.029485691338777542, + 0.04809669405221939, + 0.023906070739030838, + 0.021518472582101822, + 0.029764866456389427, + 0.03991532325744629, + 0.03102874755859375, + 0.030303189530968666, + 0.03678075969219208, + 0.030173664912581444, + 0.044436074793338776, + 0.03792708367109299, + 0.027041960507631302, + 0.021432282403111458, + 0.001865062746219337, + 0.0016597835347056389, + 0.003720928682014346, + 0.002899150364100933, + 0.001929426216520369, + 0.004753903951495886, + 0.0024423301219940186, + 0.0038332862313836813, + 0.006001933477818966, + 0.007437508087605238, + 0.009909773245453835, + 0.019872477278113365, + 0.006995126139372587, + 0.007507889531552792, + 0.012339329347014427, + 0.0093922670930624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.18369494378566742 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.225624218583107 + ], + "opposite_logprob_tokens": [ + -1.6006242036819458 + ], + "logprob_reward": -0.225624218583107, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.225624218583107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06530199199914932, + 0.03609587997198105, + 0.034659434109926224, + 0.03711716830730438, + 0.13748915493488312, + 0.08133818954229355, + 0.10115838795900345, + 0.33299919962882996, + 0.02595951221883297, + 0.14788106083869934 + ], + "credit_map": [ + 0.0057657682336866856, + 0.005634769331663847, + 0.010162207297980785, + 0.006072914693504572, + 0.007085321471095085, + 0.008552311919629574, + 0.009136546403169632, + 0.01289215125143528, + 0.0033661338966339827, + 0.004150587599724531, + 0.005823932122439146, + 0.006002354901283979, + 0.004640829749405384, + 0.0043298169039189816, + 0.003583670826628804, + 0.00419855210930109, + 0.003300476586446166, + 0.0029155409429222345, + 0.004584736656397581, + 0.003973209764808416, + 0.0051412456668913364, + 0.0050835688598454, + 0.005530937574803829, + 0.004129720386117697, + 0.0028594655450433493, + 0.0035107270814478397, + 0.004344646818935871, + 0.004591871984302998, + 0.005906571168452501, + 0.0039648073725402355, + 0.0070411041378974915, + 0.004897976294159889, + 0.010043761692941189, + 0.014231155626475811, + 0.01849948614835739, + 0.02002149634063244, + 0.019321179017424583, + 0.016016146168112755, + 0.02070198766887188, + 0.01865394599735737, + 0.008038950152695179, + 0.009575649164617062, + 0.010927315801382065, + 0.010526951402425766, + 0.009698842652142048, + 0.009807504713535309, + 0.010578103363513947, + 0.012184878811240196, + 0.008304170332849026, + 0.010411289520561695, + 0.01301867701113224, + 0.01952529512345791, + 0.01066905539482832, + 0.008365465328097343, + 0.01214636117219925, + 0.018718073144555092, + 0.057898420840501785, + 0.03043319098651409, + 0.062283243983983994, + 0.04021941497921944, + 0.03075334243476391, + 0.05707888677716255, + 0.026964202523231506, + 0.027368472889065742, + 0.002137082628905773, + 0.0024996737483888865, + 0.003773842006921768, + 0.004038968123495579, + 0.002543634967878461, + 0.0035726141650229692, + 0.003317286493256688, + 0.004076410550624132, + 0.010812869295477867, + 0.0170962642878294, + 0.021632449701428413, + 0.033494677394628525, + 0.009620696306228638, + 0.018912585452198982, + 0.020180922001600266, + 0.0161306019872427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.225624218583107 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036201435141265392 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.0036201435141265392, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036201435141265392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22181077301502228, + 0.07416342198848724, + 0.06658840924501419, + 0.026000279933214188, + 0.1358923465013504, + 0.07291492074728012, + 0.12725798785686493, + 0.16783611476421356, + 0.043866872787475586, + 0.06366892904043198 + ], + "credit_map": [ + 0.018142515793442726, + 0.016760732978582382, + 0.04551665112376213, + 0.02391708828508854, + 0.021738167852163315, + 0.02099006064236164, + 0.020152738317847252, + 0.0545928031206131, + 0.006350256502628326, + 0.00861743651330471, + 0.00952092930674553, + 0.01216110773384571, + 0.009155790321528912, + 0.00826936960220337, + 0.011095461435616016, + 0.008993071503937244, + 0.0056105065159499645, + 0.006224117241799831, + 0.010036662220954895, + 0.007308670319616795, + 0.007731372024863958, + 0.011119872331619263, + 0.008629621006548405, + 0.009927583858370781, + 0.0034141605719923973, + 0.0025228350423276424, + 0.0036449709441512823, + 0.003504134714603424, + 0.0028086763340979815, + 0.003543499391525984, + 0.0036355385091155767, + 0.0029264630284160376, + 0.012650083750486374, + 0.01905471459031105, + 0.019333116710186005, + 0.02430756948888302, + 0.013868225738406181, + 0.018844768404960632, + 0.014637275598943233, + 0.013196585699915886, + 0.00785333663225174, + 0.007013036869466305, + 0.00933061446994543, + 0.010467616841197014, + 0.008427618071436882, + 0.007748788222670555, + 0.015275394544005394, + 0.00679850485175848, + 0.0113470908254385, + 0.013293817639350891, + 0.015282934531569481, + 0.020114418119192123, + 0.01766001246869564, + 0.01251253392547369, + 0.018450750038027763, + 0.018596427515149117, + 0.025218311697244644, + 0.01988093927502632, + 0.02574627660214901, + 0.017278660088777542, + 0.020739616826176643, + 0.023900166153907776, + 0.013696687296032906, + 0.021375441923737526, + 0.0051052020862698555, + 0.00506791053339839, + 0.00603356072679162, + 0.005066639743745327, + 0.006175481714308262, + 0.006469327490776777, + 0.004369927104562521, + 0.005578824318945408, + 0.005407833959907293, + 0.005784294568002224, + 0.008687236346304417, + 0.013532334007322788, + 0.005249406676739454, + 0.006015809718519449, + 0.01142751332372427, + 0.007564490661025047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0036201435141265392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820206806063652 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820206806063652, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820206806063652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11102452129125595, + 0.05946025997400284, + 0.05060846731066704, + 0.06176866218447685, + 0.09072685986757278, + 0.13476790487766266, + 0.09992888569831848, + 0.10979233682155609, + 0.035245802253484726, + 0.2466762810945511 + ], + "credit_map": [ + 0.011196193285286427, + 0.011808454059064388, + 0.01473631989210844, + 0.010286595672369003, + 0.015331976115703583, + 0.013796842657029629, + 0.013864179141819477, + 0.020003963261842728, + 0.006841612048447132, + 0.005474936682730913, + 0.007364725694060326, + 0.009356575086712837, + 0.007246959488838911, + 0.0076272510923445225, + 0.008666247129440308, + 0.006881953217089176, + 0.004513805266469717, + 0.005173790734261274, + 0.00687636760994792, + 0.005957894027233124, + 0.007917509414255619, + 0.00695103220641613, + 0.007209821604192257, + 0.00600825110450387, + 0.007662126328796148, + 0.004719107411801815, + 0.008503030985593796, + 0.007886799052357674, + 0.006658568046987057, + 0.01191248744726181, + 0.00664192158728838, + 0.0077846166677773, + 0.008533290587365627, + 0.0109639260917902, + 0.011398357339203358, + 0.012583228759467602, + 0.013405393809080124, + 0.011241819709539413, + 0.013515299186110497, + 0.009085538797080517, + 0.012950575910508633, + 0.01603722758591175, + 0.01826976239681244, + 0.014577727764844894, + 0.014317462220788002, + 0.0177154541015625, + 0.02174033783376217, + 0.0191593449562788, + 0.009102510288357735, + 0.009118225425481796, + 0.01332005299627781, + 0.016715319827198982, + 0.011690078303217888, + 0.008571755141019821, + 0.013463718816637993, + 0.01794721931219101, + 0.014577814377844334, + 0.014687390066683292, + 0.014180128462612629, + 0.012339902110397816, + 0.017325211316347122, + 0.013104566372931004, + 0.00989602506160736, + 0.013681295327842236, + 0.003990852739661932, + 0.003876606235280633, + 0.004830001853406429, + 0.004146542400121689, + 0.004486478865146637, + 0.005941710900515318, + 0.0034343169536441565, + 0.004539292771369219, + 0.018012601882219315, + 0.027909236028790474, + 0.03775769844651222, + 0.05424991622567177, + 0.01734037697315216, + 0.03212372958660126, + 0.03193887695670128, + 0.027343858033418655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.002820206806063652 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot vertically from above in a drone perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera descending slowly and steadily throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central object a weathered, square wooden box with dark, damp, and cracked wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the box feature rusty metal bands and scattered green algae on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the box rising and falling rhythmically with gentle waves while leaving a faint wake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818355917930603 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.01818355917930603, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818355917930603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11052834242582321, + 0.04034212976694107, + 0.03222321346402168, + 0.08058322966098785, + 0.03808354586362839, + 0.06343188881874084, + 0.45214399695396423, + 0.10044818371534348, + 0.03475490212440491, + 0.047460541129112244 + ], + "credit_map": [ + 0.01609966903924942, + 0.008175911381840706, + 0.01879587396979332, + 0.008724093437194824, + 0.01000860147178173, + 0.013198877684772015, + 0.009661687538027763, + 0.025863641873002052, + 0.005432216450572014, + 0.0039569661021232605, + 0.00605193804949522, + 0.004484320525079966, + 0.005791607778519392, + 0.00636852253228426, + 0.0041727349162101746, + 0.004083828069269657, + 0.0028186545241624117, + 0.003930457402020693, + 0.0038652687799185514, + 0.0032827502582222223, + 0.003996513318270445, + 0.005776275414973497, + 0.004050697200000286, + 0.004502600524574518, + 0.00975320115685463, + 0.006106534507125616, + 0.010554813779890537, + 0.012914864346385002, + 0.008236760273575783, + 0.01357428077608347, + 0.009379111230373383, + 0.010063670575618744, + 0.003449441632255912, + 0.003952035214751959, + 0.006442153826355934, + 0.005248807370662689, + 0.004961848724633455, + 0.004441817291080952, + 0.004754276014864445, + 0.004833169747143984, + 0.006168936844915152, + 0.0066718109883368015, + 0.007897762581706047, + 0.007672660052776337, + 0.007781144231557846, + 0.009377180598676205, + 0.008896458894014359, + 0.008965939283370972, + 0.03766642138361931, + 0.04130749776959419, + 0.05831912159919739, + 0.08704457432031631, + 0.04680918529629707, + 0.038112252950668335, + 0.05747894570231438, + 0.08540602028369904, + 0.017584295943379402, + 0.009268797002732754, + 0.018832005560398102, + 0.012072713114321232, + 0.009049738757312298, + 0.01722257025539875, + 0.00786518957465887, + 0.008552881889045238, + 0.00353448954410851, + 0.003756060032173991, + 0.004752684850245714, + 0.004648657515645027, + 0.004151215311139822, + 0.005490440875291824, + 0.00439182436093688, + 0.004029531497508287, + 0.0032892886083573103, + 0.004505028948187828, + 0.007002233061939478, + 0.008347617462277412, + 0.0039589195512235165, + 0.00832416769117117, + 0.005223365966230631, + 0.006809922866523266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.01818355917930603 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_10/index_tensor([9484, 9484])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_10/index_tensor([9484, 9484])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ba036880aed715a55bef948b3c39f0bc010f8b73 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_10/index_tensor([9484, 9484])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, handheld, medium shot panning rightward plunges the viewer into the depths of a deep, dark cave. At the center of the frame is an explorer with tousled dark brown hair and a face stained with dirt. His expression is focused and alert. He wears a frayed dark green canvas jacket, layered over a gray Henley shirt, its surface darkened by the cave's dampness. In his left hand, he holds a burning torch. The dancing flame, the only source of light in the pitch black, crackles continuously and casts a warm, dynamic orange-yellow glow on the surroundings. This unstable light casts swaying, elongated shadows on the damp, craggy rock face, covered in shimmering moss and dripping with crystal-clear water droplets. He moves cautiously along the rock face, leaning slightly forward, his right hand extended, his fingertips gently and cautiously touching the slippery surface. His gaze follows the movement of his finger, scanning for any crack or unusual protrusion that might hold a clue. The entire scene creates an atmosphere of tension, mystery, and exploration with dynamic lighting, focused movements, and the crackling of torches.", + "seed": 7051, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_10/index_9484_seed_7051_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07545093726366758, + "vlm_logprob": -0.07545093726366758, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.07545093726366758, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26900365948677063 + ], + "opposite_logprob_tokens": [ + -1.5190036296844482 + ], + "logprob_reward": -0.26900365948677063, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.26900365948677063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10263792425394058, + 0.05944113805890083, + 0.10569032281637192, + 0.13552910089492798, + 0.048016685992479324, + 0.059018298983573914, + 0.060004908591508865, + 0.061244845390319824, + 0.20359134674072266, + 0.164825439453125 + ], + "credit_map": [ + 0.005904252175241709, + 0.012647494673728943, + 0.010534885339438915, + 0.01447607297450304, + 0.02744164504110813, + 0.007497571408748627, + 0.014307823032140732, + 0.009828180074691772, + 0.006901731248944998, + 0.006549553945660591, + 0.01008088979870081, + 0.007681993301957846, + 0.006271871738135815, + 0.005791239906102419, + 0.008909597992897034, + 0.007254256866872311, + 0.006173540838062763, + 0.030224688351154327, + 0.009955315850675106, + 0.013266191817820072, + 0.015195384621620178, + 0.009138388559222221, + 0.012314834631979465, + 0.00942198745906353, + 0.012338392436504364, + 0.032734718173742294, + 0.014688820578157902, + 0.01587854139506817, + 0.014495305716991425, + 0.010093518532812595, + 0.01960759051144123, + 0.0156922098249197, + 0.007375185377895832, + 0.012363806366920471, + 0.006669167894870043, + 0.0057455794885754585, + 0.002898103091865778, + 0.004018391482532024, + 0.0035799178294837475, + 0.005366533994674683, + 0.008275488391518593, + 0.00704730860888958, + 0.0065408810041844845, + 0.0069167837500572205, + 0.007490148302167654, + 0.005986223462969065, + 0.006884214002639055, + 0.009877251461148262, + 0.006445613224059343, + 0.005039261654019356, + 0.007984565570950508, + 0.007373193744570017, + 0.010158788412809372, + 0.004883554298430681, + 0.009405657649040222, + 0.008714267984032631, + 0.006613923702389002, + 0.009521625936031342, + 0.008046519942581654, + 0.011759692803025246, + 0.006117017474025488, + 0.006613084115087986, + 0.005275625269860029, + 0.007297361269593239, + 0.02674427255988121, + 0.02647131122648716, + 0.02336181327700615, + 0.024224035441875458, + 0.02352885901927948, + 0.022502833977341652, + 0.03056725114583969, + 0.026190979406237602, + 0.012798367068171501, + 0.011313281022012234, + 0.031127706170082092, + 0.015767598524689674, + 0.014225245453417301, + 0.039347294718027115, + 0.02041630819439888, + 0.019829628989100456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.26900365948677063 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02984113246202469 + ], + "opposite_logprob_tokens": [ + -3.529841184616089 + ], + "logprob_reward": -0.02984113246202469, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02984113246202469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28450819849967957, + 0.2367650717496872, + 0.054776012897491455, + 0.07844546437263489, + 0.05236339569091797, + 0.043882448226213455, + 0.061622101813554764, + 0.09595565497875214, + 0.02654046192765236, + 0.06514128297567368 + ], + "credit_map": [ + 0.024486808106303215, + 0.038130395114421844, + 0.026743058115243912, + 0.02945277839899063, + 0.07041709870100021, + 0.025793934240937233, + 0.04684532433748245, + 0.022638767957687378, + 0.020849617198109627, + 0.01753224804997444, + 0.031717006117105484, + 0.03083026595413685, + 0.03230133280158043, + 0.03679650276899338, + 0.04291575402021408, + 0.023822307586669922, + 0.004720936994999647, + 0.013429220765829086, + 0.006386646069586277, + 0.005397514905780554, + 0.006764792371541262, + 0.004871901124715805, + 0.0073204138316214085, + 0.005884586367756128, + 0.006323121953755617, + 0.011906673200428486, + 0.014839423820376396, + 0.008663451299071312, + 0.008481746539473534, + 0.009717710316181183, + 0.009588263928890228, + 0.008925065398216248, + 0.008818747475743294, + 0.005999924149364233, + 0.007306384854018688, + 0.00575219513848424, + 0.004925657995045185, + 0.00566856050863862, + 0.00573003850877285, + 0.008161882869899273, + 0.0048256320878863335, + 0.0048885587602853775, + 0.00793803296983242, + 0.007829242385923862, + 0.004012271296232939, + 0.0027788726147264242, + 0.004006243776530027, + 0.00760358851402998, + 0.005460766609758139, + 0.007243003696203232, + 0.008934096433222294, + 0.008962941356003284, + 0.007058095186948776, + 0.005806459579616785, + 0.009202171117067337, + 0.008954562246799469, + 0.011282221414148808, + 0.010853113606572151, + 0.014005105011165142, + 0.01852324604988098, + 0.008258518762886524, + 0.012927395291626453, + 0.009340200573205948, + 0.01076585240662098, + 0.002499182941392064, + 0.0026885722763836384, + 0.00401297165080905, + 0.0031798796262592077, + 0.0037042549811303616, + 0.002971245674416423, + 0.003533189184963703, + 0.003951165359467268, + 0.005144910421222448, + 0.004467920865863562, + 0.01801750622689724, + 0.00495132664218545, + 0.005704677663743496, + 0.016564272344112396, + 0.005009263753890991, + 0.00528140040114522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02984113246202469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038118183612823486 + ], + "opposite_logprob_tokens": [ + -3.2881181240081787 + ], + "logprob_reward": -0.038118183612823486, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038118183612823486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19545184075832367, + 0.0935031846165657, + 0.1308598667383194, + 0.1477992832660675, + 0.08516492694616318, + 0.07146658003330231, + 0.058114681392908096, + 0.12076183408498764, + 0.02078806795179844, + 0.07608968764543533 + ], + "credit_map": [ + 0.013290079310536385, + 0.019443310797214508, + 0.039864249527454376, + 0.02165778912603855, + 0.0220181941986084, + 0.02230042777955532, + 0.03506281226873398, + 0.02181500382721424, + 0.009819183498620987, + 0.006910127587616444, + 0.022855399176478386, + 0.007503715343773365, + 0.012552890926599503, + 0.014710651710629463, + 0.009856878779828548, + 0.009294346906244755, + 0.026625538244843483, + 0.01760469190776348, + 0.014774304814636707, + 0.0108408834785223, + 0.016099151223897934, + 0.013062111102044582, + 0.018347911536693573, + 0.013505284674465656, + 0.014496653340756893, + 0.03520382195711136, + 0.015593431890010834, + 0.016423078253865242, + 0.014778465032577515, + 0.010760776698589325, + 0.023504719138145447, + 0.01703835278749466, + 0.010814180597662926, + 0.013169207610189915, + 0.014936898835003376, + 0.012408783659338951, + 0.011113687418401241, + 0.005436117760837078, + 0.008899843320250511, + 0.008386215195059776, + 0.008540068753063679, + 0.010964359156787395, + 0.008555515669286251, + 0.007136018481105566, + 0.011066650971770287, + 0.009067321196198463, + 0.006065257824957371, + 0.01007139217108488, + 0.006620707456022501, + 0.005675519350916147, + 0.0070359171368181705, + 0.007704723626375198, + 0.008232900872826576, + 0.004840941168367863, + 0.009745722636580467, + 0.008258254267275333, + 0.013827992603182793, + 0.014910793863236904, + 0.017751367762684822, + 0.017087625339627266, + 0.013189730234444141, + 0.01896381750702858, + 0.011227572336792946, + 0.013802934437990189, + 0.0027507441118359566, + 0.0016897302120923996, + 0.0032278550788760185, + 0.0021680807694792747, + 0.0023748178500682116, + 0.002473911503329873, + 0.002941644750535488, + 0.0031612846069037914, + 0.005964161362498999, + 0.007350486237555742, + 0.016461875289678574, + 0.006690467707812786, + 0.0071695726364851, + 0.01896568201482296, + 0.00684047257527709, + 0.006646974477916956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.038118183612823486 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198542468249798 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198542468249798, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198542468249798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08776605129241943, + 0.06478646397590637, + 0.05585649237036705, + 0.1939520388841629, + 0.06009063497185707, + 0.11536210030317307, + 0.13218186795711517, + 0.07629194110631943, + 0.11186663061380386, + 0.10184582322835922 + ], + "credit_map": [ + 0.00821873638778925, + 0.008930398151278496, + 0.013475917279720306, + 0.010076349601149559, + 0.011168858036398888, + 0.012734664604067802, + 0.013444316573441029, + 0.009716801345348358, + 0.007480884436517954, + 0.006427271291613579, + 0.01713748462498188, + 0.006076537538319826, + 0.007981179282069206, + 0.006134111434221268, + 0.00824555940926075, + 0.0053034317679703236, + 0.006484359502792358, + 0.011157604865729809, + 0.005556394811719656, + 0.006454981863498688, + 0.007539014331996441, + 0.006182917859405279, + 0.007476252503693104, + 0.005004962906241417, + 0.01752283051609993, + 0.034373100847005844, + 0.021181980147957802, + 0.019467052072286606, + 0.0297886673361063, + 0.023256082087755203, + 0.023688117042183876, + 0.024674201384186745, + 0.008221284486353397, + 0.00858366396278143, + 0.00898057036101818, + 0.008421434089541435, + 0.00464077852666378, + 0.005867403466254473, + 0.0046480437740683556, + 0.010727452114224434, + 0.011956554837524891, + 0.01309126801788807, + 0.015828263014554977, + 0.020995663478970528, + 0.013693689368665218, + 0.012507033534348011, + 0.01093575730919838, + 0.01635386049747467, + 0.014150169678032398, + 0.01038078311830759, + 0.01386498473584652, + 0.016453580930829048, + 0.02496146224439144, + 0.011420484632253647, + 0.022760964930057526, + 0.01818942278623581, + 0.007599284872412682, + 0.00842609815299511, + 0.01024994533509016, + 0.008617289364337921, + 0.0055529591627418995, + 0.015321592800319195, + 0.007691207807511091, + 0.012833555229008198, + 0.01465389970690012, + 0.014979693107306957, + 0.013496814295649529, + 0.015652917325496674, + 0.014113375917077065, + 0.014369955286383629, + 0.013137816451489925, + 0.011462155729532242, + 0.01073397696018219, + 0.009218201972544193, + 0.01638871617615223, + 0.011818153783679008, + 0.013961040414869785, + 0.014553185552358627, + 0.0149935781955719, + 0.010178964585065842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.002198542468249798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038093168288469315 + ], + "opposite_logprob_tokens": [ + -3.288093090057373 + ], + "logprob_reward": -0.038093168288469315, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038093168288469315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1357031762599945, + 0.0846189484000206, + 0.07952065765857697, + 0.13082611560821533, + 0.07070723176002502, + 0.08246899396181107, + 0.08870915323495865, + 0.16619892418384552, + 0.07460758090019226, + 0.08663925528526306 + ], + "credit_map": [ + 0.012015385553240776, + 0.014717352576553822, + 0.01583024673163891, + 0.021428877487778664, + 0.017968902364373207, + 0.020879074931144714, + 0.019112607464194298, + 0.013750731013715267, + 0.008558284491300583, + 0.008119828999042511, + 0.012716119177639484, + 0.00931699201464653, + 0.018564647063612938, + 0.008031804114580154, + 0.011317710392177105, + 0.007993564009666443, + 0.008995059877634048, + 0.011771167628467083, + 0.009700113907456398, + 0.006669771391898394, + 0.017078103497624397, + 0.008162636309862137, + 0.010241583921015263, + 0.006902218796312809, + 0.011784563772380352, + 0.03256167843937874, + 0.013642411679029465, + 0.014443456195294857, + 0.016392821446061134, + 0.010196048766374588, + 0.01594569720327854, + 0.015859439969062805, + 0.010608658194541931, + 0.009248627349734306, + 0.009468726813793182, + 0.0066233729012310505, + 0.005484370049089193, + 0.010478492826223373, + 0.006402661558240652, + 0.012392319738864899, + 0.009188639931380749, + 0.00957789458334446, + 0.012735159136354923, + 0.008559891022741795, + 0.012164655141532421, + 0.010204882360994816, + 0.00846592616289854, + 0.011571944691240788, + 0.010551048442721367, + 0.0049411337822675705, + 0.009863420389592648, + 0.0071427831426262856, + 0.015512709505856037, + 0.006209506653249264, + 0.016550850123167038, + 0.01793770119547844, + 0.022474370896816254, + 0.015114824287593365, + 0.030095715075731277, + 0.023177191615104675, + 0.016009967774152756, + 0.02694152295589447, + 0.015176087617874146, + 0.0172092467546463, + 0.006308701355010271, + 0.007582616060972214, + 0.011484264396131039, + 0.008692326955497265, + 0.00998055562376976, + 0.008242842741310596, + 0.011024805717170238, + 0.011291464790701866, + 0.012584660202264786, + 0.009969566017389297, + 0.010198161005973816, + 0.010135638527572155, + 0.013869203627109528, + 0.01009317021816969, + 0.01055154763162136, + 0.009237309917807579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.038093168288469315 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, handheld, medium shot panning rightward plunges the viewer into the depths of a deep, dark cave. At the center of the frame is an explorer with tousled dark brown hair and a face stained with dirt. His expression is focused and alert. He wears a frayed dark green canvas jacket, layered over a gray Henley shirt, its surface darkened by the cave's dampness. In his left hand, he holds a burning torch. The dancing flame, the only source of light in the pitch black, crackles continuously and casts a warm, dynamic orange-yellow glow on the surroundings. This unstable light casts swaying, elongated shadows on the damp, craggy rock face, covered in shimmering moss and dripping with crystal-clear water droplets. He moves cautiously along the rock face, leaning slightly forward, his right hand extended, his fingertips gently and cautiously touching the slippery surface. His gaze follows the movement of his finger, scanning for any crack or unusual protrusion that might hold a clue. The entire scene creates an atmosphere of tension, mystery, and exploration with dynamic lighting, focused movements, and the crackling of torches.", + "seed": 7052, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_10/index_9484_seed_7052_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.057492156000807884, + "vlm_logprob": -0.057492156000807884, + "vlm_yes_no_margin": 3.600000023841858, + "vlm_token_credit": -0.057492156000807884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15820659697055817 + ], + "opposite_logprob_tokens": [ + -2.0332067012786865 + ], + "logprob_reward": -0.15820659697055817, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.15820659697055817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24074342846870422, + 0.06810598820447922, + 0.06633112579584122, + 0.08616737276315689, + 0.19757503271102905, + 0.17888198792934418, + 0.0841270238161087, + 0.03563601151108742, + 0.016006316989660263, + 0.02642565220594406 + ], + "credit_map": [ + 0.019603043794631958, + 0.030026480555534363, + 0.010903958231210709, + 0.014917991124093533, + 0.01655040867626667, + 0.0701313316822052, + 0.01607033982872963, + 0.06253989785909653, + 0.005328363738954067, + 0.00880719255656004, + 0.0048775807954370975, + 0.008048239164054394, + 0.012110674753785133, + 0.01493373978883028, + 0.007343180477619171, + 0.006657016929239035, + 0.004915803205221891, + 0.005295462440699339, + 0.006052141077816486, + 0.0060774474404752254, + 0.013439737260341644, + 0.013693531043827534, + 0.008704871870577335, + 0.00815214216709137, + 0.008398598060011864, + 0.008739745244383812, + 0.01216756273061037, + 0.009978157468140125, + 0.00863424502313137, + 0.016963519155979156, + 0.008811533451080322, + 0.012474020011723042, + 0.01977211982011795, + 0.01648503728210926, + 0.025199446827173233, + 0.028963878750801086, + 0.01625298149883747, + 0.03052089922130108, + 0.025645863264799118, + 0.03473483771085739, + 0.01933060772716999, + 0.029609767720103264, + 0.026617785915732384, + 0.02809191681444645, + 0.016505088657140732, + 0.023714274168014526, + 0.018567441031336784, + 0.01644512079656124, + 0.00865865033119917, + 0.010778790339827538, + 0.015523137524724007, + 0.007640241179615259, + 0.00812347512692213, + 0.0083067137748003, + 0.01385466754436493, + 0.011241355910897255, + 0.0031319677364081144, + 0.004001887049525976, + 0.005314819049090147, + 0.0033702398650348186, + 0.0034789626952260733, + 0.0028871181420981884, + 0.009169720113277435, + 0.004281302448362112, + 0.0013022377388551831, + 0.0017074199859052896, + 0.004546143114566803, + 0.0023758788593113422, + 0.0013676100643351674, + 0.0011407452402636409, + 0.002617361955344677, + 0.0009489226504229009, + 0.0026130806654691696, + 0.003985777497291565, + 0.004118261858820915, + 0.004562036599963903, + 0.002276028273627162, + 0.003900188021361828, + 0.002282055327668786, + 0.002688224194571376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.15820659697055817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07905685901641846 + ], + "opposite_logprob_tokens": [ + -2.579056739807129 + ], + "logprob_reward": -0.07905685901641846, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905685901641846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2903416156768799, + 0.109678715467453, + 0.12843038141727448, + 0.06797632575035095, + 0.09099175781011581, + 0.1398485004901886, + 0.08407802879810333, + 0.029350463300943375, + 0.03372825309634209, + 0.02557605504989624 + ], + "credit_map": [ + 0.012774134054780006, + 0.022590667009353638, + 0.010130567476153374, + 0.012282522395253181, + 0.022519318386912346, + 0.10796920955181122, + 0.014253135770559311, + 0.08782201260328293, + 0.008338365703821182, + 0.010873420163989067, + 0.006318571045994759, + 0.011908164247870445, + 0.02560860477387905, + 0.021187109872698784, + 0.013245617039501667, + 0.012198838405311108, + 0.010115365497767925, + 0.018907852470874786, + 0.00814083218574524, + 0.011279304511845112, + 0.0184621699154377, + 0.01827242411673069, + 0.02978944033384323, + 0.013462982140481472, + 0.005076896399259567, + 0.009761708788573742, + 0.007279494311660528, + 0.01079858560115099, + 0.009741523303091526, + 0.007811808958649635, + 0.007135201245546341, + 0.010371094569563866, + 0.008661557920277119, + 0.01109985075891018, + 0.011042761616408825, + 0.01112364511936903, + 0.00828480813652277, + 0.014015157707035542, + 0.01283306535333395, + 0.01393089722841978, + 0.01299871876835823, + 0.021536996588110924, + 0.017835652455687523, + 0.022079937160015106, + 0.01076981145888567, + 0.018820034340023994, + 0.02455848827958107, + 0.011248838156461716, + 0.007973067462444305, + 0.0097262067720294, + 0.01285190787166357, + 0.00850907526910305, + 0.008275614120066166, + 0.007443694397807121, + 0.021176693961024284, + 0.008121749386191368, + 0.0029939482919871807, + 0.003320265095680952, + 0.004277219530194998, + 0.003077981062233448, + 0.002255222527310252, + 0.0017105182632803917, + 0.006450315937399864, + 0.005264986772090197, + 0.0026030917651951313, + 0.005992100574076176, + 0.006845096591860056, + 0.004936970770359039, + 0.003049413440749049, + 0.002551499754190445, + 0.005601454060524702, + 0.0021486186888068914, + 0.0027103836182504892, + 0.005319884046912193, + 0.0036845807917416096, + 0.0032980756368488073, + 0.0021981524769216776, + 0.003413355676457286, + 0.0024549076333642006, + 0.002496710978448391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.07905685901641846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201468512415886 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201468512415886, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201468512415886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23440663516521454, + 0.051541201770305634, + 0.23563791811466217, + 0.1642080694437027, + 0.1262694001197815, + 0.07219251245260239, + 0.04133988171815872, + 0.020542019978165627, + 0.01964116096496582, + 0.034221235662698746 + ], + "credit_map": [ + 0.011275960132479668, + 0.014780707657337189, + 0.007169025484472513, + 0.009258621372282505, + 0.017417404800653458, + 0.09281375259160995, + 0.01015479862689972, + 0.0715363547205925, + 0.004599712323397398, + 0.005689399316906929, + 0.002999483607709408, + 0.008345584385097027, + 0.0054783509112894535, + 0.010065537877380848, + 0.008991076610982418, + 0.005372055806219578, + 0.016380049288272858, + 0.04538039490580559, + 0.02479139342904091, + 0.026959428563714027, + 0.04124575853347778, + 0.03278248384594917, + 0.029255930334329605, + 0.018842466175556183, + 0.014691067859530449, + 0.025065720081329346, + 0.019403154030442238, + 0.025159599259495735, + 0.015215152874588966, + 0.01950017362833023, + 0.01545204408466816, + 0.029721152037382126, + 0.013288293033838272, + 0.0155388955026865, + 0.016069795936346054, + 0.015643620863556862, + 0.011097487062215805, + 0.02127722278237343, + 0.014290176331996918, + 0.0190639179199934, + 0.007147175259888172, + 0.008155724965035915, + 0.009656351059675217, + 0.01242933701723814, + 0.005179851781576872, + 0.012640267610549927, + 0.008468445390462875, + 0.00851536076515913, + 0.0032680470030754805, + 0.0049128406681120396, + 0.008225388824939728, + 0.005771159660071135, + 0.0038203285075724125, + 0.0040441821329295635, + 0.007570818532258272, + 0.0037271182518452406, + 0.002005757763981819, + 0.001773951342329383, + 0.0027097687125205994, + 0.002176014008000493, + 0.002274252474308014, + 0.0013222347479313612, + 0.00474072340875864, + 0.0035393182188272476, + 0.0015029179630801082, + 0.003344126744195819, + 0.005121838301420212, + 0.0022321995347738266, + 0.0017929567256942391, + 0.0015839304542168975, + 0.0029183062724769115, + 0.0011448838049545884, + 0.003523011691868305, + 0.004751091357320547, + 0.007507887203246355, + 0.005054791457951069, + 0.0029583291616290808, + 0.003860161639750004, + 0.002825183793902397, + 0.0037407781928777695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.018201468512415886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021997320000082254 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021997320000082254, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021997320000082254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24157339334487915, + 0.1001601368188858, + 0.07974442094564438, + 0.07824216037988663, + 0.11461806297302246, + 0.09068512171506882, + 0.14271730184555054, + 0.08582339435815811, + 0.02763904072344303, + 0.038796961307525635 + ], + "credit_map": [ + 0.016157986596226692, + 0.03218474239110947, + 0.008336692117154598, + 0.01378575712442398, + 0.027740992605686188, + 0.07620314508676529, + 0.018111862242221832, + 0.049052219837903976, + 0.009728945791721344, + 0.013089021667838097, + 0.005409575998783112, + 0.008354967460036278, + 0.01330834161490202, + 0.023608092218637466, + 0.016515852883458138, + 0.010145341977477074, + 0.005355676636099815, + 0.010566169396042824, + 0.007489935494959354, + 0.0077165826223790646, + 0.015947407111525536, + 0.015442312695086002, + 0.009094161912798882, + 0.008132172748446465, + 0.007586494088172913, + 0.007362814620137215, + 0.009927626699209213, + 0.010945907793939114, + 0.010936476290225983, + 0.009127459488809109, + 0.009522952139377594, + 0.012832428328692913, + 0.014735681936144829, + 0.011960433796048164, + 0.017495494335889816, + 0.013516439124941826, + 0.008093724027276039, + 0.019581234082579613, + 0.010816720314323902, + 0.018418336287140846, + 0.008746654726564884, + 0.010055079124867916, + 0.011380265466868877, + 0.015611077658832073, + 0.007317266892641783, + 0.013967420905828476, + 0.01523459143936634, + 0.008372766897082329, + 0.011924373917281628, + 0.015365703031420708, + 0.021084854379296303, + 0.016052350401878357, + 0.01623713970184326, + 0.019698375836014748, + 0.030805297195911407, + 0.011549211107194424, + 0.008206695318222046, + 0.00558004155755043, + 0.011918993666768074, + 0.009473763406276703, + 0.006218763999640942, + 0.0045758504420518875, + 0.022816579788923264, + 0.01703270524740219, + 0.0026222756132483482, + 0.004660810809582472, + 0.004329856019467115, + 0.004242558963596821, + 0.003507924033328891, + 0.0032335654832422733, + 0.002931720344349742, + 0.0021103299222886562, + 0.003724419279024005, + 0.006715159863233566, + 0.006251602899283171, + 0.004829694516956806, + 0.00438026012852788, + 0.005041557364165783, + 0.003650929080322385, + 0.004203337244689465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.0021997320000082254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796123504638672 + ], + "opposite_logprob_tokens": [ + -3.5297961235046387 + ], + "logprob_reward": -0.029796123504638672, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796123504638672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2654181718826294, + 0.10765498131513596, + 0.06831509619951248, + 0.03549954667687416, + 0.242205411195755, + 0.14724180102348328, + 0.03189535066485405, + 0.04154205694794655, + 0.018516896292567253, + 0.0417107455432415 + ], + "credit_map": [ + 0.017244962975382805, + 0.02968726120889187, + 0.00826556421816349, + 0.014255008660256863, + 0.022983131930232048, + 0.08398686349391937, + 0.023839427158236504, + 0.06515592336654663, + 0.00971636176109314, + 0.008650146424770355, + 0.003357842331752181, + 0.007160588633269072, + 0.029817113652825356, + 0.027703041210770607, + 0.011246347799897194, + 0.010003529489040375, + 0.004338513594120741, + 0.00889597088098526, + 0.008177918381989002, + 0.008552560582756996, + 0.010876690968871117, + 0.01265672966837883, + 0.00737843569368124, + 0.007438264787197113, + 0.003745729336515069, + 0.0033402531407773495, + 0.0047592585906386375, + 0.005252573173493147, + 0.004503529518842697, + 0.005572761874645948, + 0.0035310080274939537, + 0.004794431384652853, + 0.03103563003242016, + 0.0250566266477108, + 0.04109816253185272, + 0.026516351848840714, + 0.017709877341985703, + 0.05144045129418373, + 0.023685285821557045, + 0.025662990286946297, + 0.012882129289209843, + 0.016661489382386208, + 0.01651647500693798, + 0.025427961722016335, + 0.012403111904859543, + 0.02051888406276703, + 0.029996253550052643, + 0.01283547654747963, + 0.0032519116066396236, + 0.004554936196655035, + 0.005014811176806688, + 0.0029691222589462996, + 0.0034443975891917944, + 0.0043279752135276794, + 0.004880995489656925, + 0.0034511990379542112, + 0.004145421553403139, + 0.004323828499764204, + 0.005704974289983511, + 0.0034491403494030237, + 0.0027816547080874443, + 0.0028987624682486057, + 0.010129979811608791, + 0.008108289912343025, + 0.001711472636088729, + 0.004047268535941839, + 0.0023510297760367393, + 0.0028149220161139965, + 0.0018394454382359982, + 0.002502621617168188, + 0.002009416464716196, + 0.001240718993358314, + 0.0038768413942307234, + 0.006797553971409798, + 0.00846290122717619, + 0.004414621274918318, + 0.0034812274388968945, + 0.00578865222632885, + 0.0032937165815383196, + 0.0055952309630811214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.029796123504638672 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, handheld, medium shot panning rightward plunges the viewer into the depths of a deep, dark cave. At the center of the frame is an explorer with tousled dark brown hair and a face stained with dirt. His expression is focused and alert. He wears a frayed dark green canvas jacket, layered over a gray Henley shirt, its surface darkened by the cave's dampness. In his left hand, he holds a burning torch. The dancing flame, the only source of light in the pitch black, crackles continuously and casts a warm, dynamic orange-yellow glow on the surroundings. This unstable light casts swaying, elongated shadows on the damp, craggy rock face, covered in shimmering moss and dripping with crystal-clear water droplets. He moves cautiously along the rock face, leaning slightly forward, his right hand extended, his fingertips gently and cautiously touching the slippery surface. His gaze follows the movement of his finger, scanning for any crack or unusual protrusion that might hold a clue. The entire scene creates an atmosphere of tension, mystery, and exploration with dynamic lighting, focused movements, and the crackling of torches.", + "seed": 7053, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_10/index_9484_seed_7053_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08703870181925595, + "vlm_logprob": -0.08703870181925595, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.08703870181925595, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3277929127216339 + ], + "opposite_logprob_tokens": [ + -1.3277928829193115 + ], + "logprob_reward": -0.3277929127216339, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3277929127216339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10949286818504333, + 0.06828026473522186, + 0.11846812069416046, + 0.0855356901884079, + 0.12441718578338623, + 0.10749832540750504, + 0.07793799042701721, + 0.03020276129245758, + 0.07798115164041519, + 0.200185626745224 + ], + "credit_map": [ + 0.018084531649947166, + 0.008228737860918045, + 0.019382212311029434, + 0.009798554703593254, + 0.016501272097229958, + 0.00714236032217741, + 0.01747295632958412, + 0.012882241047918797, + 0.008918062783777714, + 0.00676931394264102, + 0.010246340185403824, + 0.007798958104103804, + 0.007323661353439093, + 0.00526683172211051, + 0.009003642946481705, + 0.012953449040651321, + 0.014498792588710785, + 0.015150398947298527, + 0.008860943838953972, + 0.01589890941977501, + 0.01963963732123375, + 0.010636227205395699, + 0.015215542167425156, + 0.018567660823464394, + 0.016043489798903465, + 0.014180661179125309, + 0.011414525099098682, + 0.0062143560498952866, + 0.008834159933030605, + 0.007670274935662746, + 0.012136625126004219, + 0.00904159341007471, + 0.013414640910923481, + 0.018381010740995407, + 0.013248933479189873, + 0.014784987084567547, + 0.013718202710151672, + 0.020828232169151306, + 0.015291111543774605, + 0.014750066213309765, + 0.007900429889559746, + 0.016280505806207657, + 0.012055695056915283, + 0.013415112160146236, + 0.022034205496311188, + 0.01147492602467537, + 0.009278320707380772, + 0.015059120021760464, + 0.006036411505192518, + 0.004306136630475521, + 0.0073335617780685425, + 0.004590945318341255, + 0.006890119053423405, + 0.006801209878176451, + 0.025062991306185722, + 0.016916610300540924, + 0.0020806575194001198, + 0.00293821282684803, + 0.0034005206543952227, + 0.004689588211476803, + 0.005641639698296785, + 0.0027192067354917526, + 0.005256330594420433, + 0.003476605052128434, + 0.0065909940749406815, + 0.0054039969108998775, + 0.006372308358550072, + 0.010078021325170994, + 0.012744652107357979, + 0.007239797618240118, + 0.014391039498150349, + 0.015160338953137398, + 0.022655485197901726, + 0.018830202519893646, + 0.021909799426794052, + 0.027963196858763695, + 0.03202061355113983, + 0.02498260699212551, + 0.03034774214029312, + 0.021475976333022118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.3277929127216339 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373567387461662 + ], + "opposite_logprob_tokens": [ + -3.408735752105713 + ], + "logprob_reward": -0.03373567387461662, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373567387461662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07686507701873779, + 0.03892550989985466, + 0.10125236213207245, + 0.10915660113096237, + 0.09101134538650513, + 0.10330544412136078, + 0.11592694371938705, + 0.1406380981206894, + 0.1488424837589264, + 0.07407612353563309 + ], + "credit_map": [ + 0.011856154538691044, + 0.013644378632307053, + 0.008567352779209614, + 0.00870227999985218, + 0.009689855389297009, + 0.006284743547439575, + 0.00970181729644537, + 0.008418492041528225, + 0.005468930117785931, + 0.00592689448967576, + 0.004560532048344612, + 0.005054859910160303, + 0.004036454949527979, + 0.003142335917800665, + 0.005513072945177555, + 0.005222429055720568, + 0.011032366193830967, + 0.015886299312114716, + 0.013732454739511013, + 0.017057087272405624, + 0.010299911722540855, + 0.009690645150840282, + 0.014109872281551361, + 0.009443728253245354, + 0.01688867248594761, + 0.0134156234562397, + 0.016493409872055054, + 0.008151731453835964, + 0.010478016920387745, + 0.010265436954796314, + 0.014713195152580738, + 0.01875051110982895, + 0.010166559368371964, + 0.016263848170638084, + 0.008223513141274452, + 0.008701244369149208, + 0.008908645249903202, + 0.012016579508781433, + 0.009220696054399014, + 0.017510253936052322, + 0.009719314984977245, + 0.013396801427006721, + 0.011963293887674809, + 0.014124027453362942, + 0.020151006057858467, + 0.010611859150230885, + 0.01018432341516018, + 0.013154811225831509, + 0.012038829736411572, + 0.012239563278853893, + 0.016180945560336113, + 0.018265049904584885, + 0.014039136469364166, + 0.012668811716139317, + 0.014377549290657043, + 0.01611705869436264, + 0.006937074009329081, + 0.004050116986036301, + 0.00890522263944149, + 0.017033161595463753, + 0.007052739150822163, + 0.0046284738928079605, + 0.05342649668455124, + 0.038604818284511566, + 0.01153307594358921, + 0.00758716743439436, + 0.010449049063026905, + 0.01974634826183319, + 0.03561597317457199, + 0.020894447341561317, + 0.022015715017914772, + 0.021000714972615242, + 0.006193246226757765, + 0.005648138467222452, + 0.006016658153384924, + 0.010205628350377083, + 0.011891115456819534, + 0.012218003161251545, + 0.010655724443495274, + 0.011247609741985798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.03373567387461662 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866965487599373 + ], + "opposite_logprob_tokens": [ + -3.0486695766448975 + ], + "logprob_reward": -0.04866965487599373, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866965487599373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13483543694019318, + 0.09121067821979523, + 0.18980897963047028, + 0.09847023338079453, + 0.0870305523276329, + 0.07390234619379044, + 0.11131811887025833, + 0.04548557102680206, + 0.06415344774723053, + 0.10378467291593552 + ], + "credit_map": [ + 0.02385029010474682, + 0.02170494571328163, + 0.02664359100162983, + 0.010602359659969807, + 0.019830288365483284, + 0.007236449513584375, + 0.013822362758219242, + 0.011145145632326603, + 0.007872727699577808, + 0.013473919592797756, + 0.008130022324621677, + 0.018598634749650955, + 0.013876575045287609, + 0.009546278044581413, + 0.01055120024830103, + 0.009161325171589851, + 0.01883237436413765, + 0.03038140758872032, + 0.02116917259991169, + 0.03306281939148903, + 0.024027623236179352, + 0.017455030232667923, + 0.022616183385252953, + 0.022264372557401657, + 0.017488736659288406, + 0.013801135122776031, + 0.010747904889285564, + 0.006306328345090151, + 0.008366115391254425, + 0.01157093420624733, + 0.014989634975790977, + 0.015199446119368076, + 0.011973164975643158, + 0.01685069501399994, + 0.010839447379112244, + 0.007934963330626488, + 0.007490190211683512, + 0.011295775882899761, + 0.0102889658883214, + 0.010357343591749668, + 0.004734070040285587, + 0.011172330938279629, + 0.008418107405304909, + 0.006406685337424278, + 0.015233815647661686, + 0.011078008450567722, + 0.005416924133896828, + 0.011442404240369797, + 0.008573758415877819, + 0.008353406563401222, + 0.012067237868905067, + 0.012097730301320553, + 0.012881306000053883, + 0.010343555361032486, + 0.024887556210160255, + 0.022113559767603874, + 0.0030179154127836227, + 0.002008215757086873, + 0.003300157142803073, + 0.006649663671851158, + 0.00574202835559845, + 0.0032839972991496325, + 0.01296146959066391, + 0.008522125892341137, + 0.005634219851344824, + 0.0033846022561192513, + 0.00555185554549098, + 0.009423519484698772, + 0.012315877713263035, + 0.0077646696008741856, + 0.01143703144043684, + 0.008641676977276802, + 0.008014025166630745, + 0.009837855584919453, + 0.011220229789614677, + 0.014447011984884739, + 0.018838798627257347, + 0.01378505676984787, + 0.015158677473664284, + 0.012483025901019573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.04866965487599373 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017138330731540918 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017138330731540918, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017138330731540918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20742250978946686, + 0.058152489364147186, + 0.16044878959655762, + 0.09506535530090332, + 0.03458915278315544, + 0.066398486495018, + 0.1091264858841896, + 0.09990791976451874, + 0.07741013914346695, + 0.09147865325212479 + ], + "credit_map": [ + 0.0589209608733654, + 0.07131095230579376, + 0.013476108200848103, + 0.012772162444889545, + 0.012485155835747719, + 0.00933620985597372, + 0.02037644013762474, + 0.008744515478610992, + 0.010127972811460495, + 0.005744573660194874, + 0.011191210709512234, + 0.006179967895150185, + 0.00446179648861289, + 0.003586923237890005, + 0.010223159566521645, + 0.006636887323111296, + 0.013617994263768196, + 0.017275962978601456, + 0.019808635115623474, + 0.03165687620639801, + 0.023607194423675537, + 0.017720462754368782, + 0.020952196791768074, + 0.015809468924999237, + 0.009541762992739677, + 0.011412897147238255, + 0.013002564199268818, + 0.006045093759894371, + 0.010968090035021305, + 0.012084408663213253, + 0.014529824256896973, + 0.01748071238398552, + 0.005254319868981838, + 0.0043299561366438866, + 0.00355783523991704, + 0.0031234959606081247, + 0.0030752266757190228, + 0.006428099237382412, + 0.003617064328864217, + 0.0052031553350389, + 0.004456355702131987, + 0.008566128090023994, + 0.009779560379683971, + 0.007757620885968208, + 0.011314054019749165, + 0.00793985091149807, + 0.005928888916969299, + 0.010656030848622322, + 0.0066980221308767796, + 0.006112993229180574, + 0.011150982230901718, + 0.010216805152595043, + 0.008074137382209301, + 0.00735302921384573, + 0.030292825773358345, + 0.029227694496512413, + 0.006860412657260895, + 0.00491302041336894, + 0.007481943350285292, + 0.013296954333782196, + 0.007143191061913967, + 0.005122441332787275, + 0.0334935337305069, + 0.021596424281597137, + 0.010552543215453625, + 0.0057776146568357944, + 0.006210785824805498, + 0.009772628545761108, + 0.0070195272564888, + 0.008236547000706196, + 0.014213605783879757, + 0.01562689058482647, + 0.009186875075101852, + 0.008859527297317982, + 0.010405311360955238, + 0.012393297627568245, + 0.013474814593791962, + 0.012705964967608452, + 0.012918490916490555, + 0.011534365825355053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0017138330731540918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023281434550881386 + ], + "opposite_logprob_tokens": [ + -3.7732813358306885 + ], + "logprob_reward": -0.023281434550881386, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023281434550881386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09295875579118729, + 0.04764479026198387, + 0.3613925874233246, + 0.07765082269906998, + 0.04609145224094391, + 0.17980198562145233, + 0.03607812523841858, + 0.05849893018603325, + 0.0664849728345871, + 0.03339756652712822 + ], + "credit_map": [ + 0.025151662528514862, + 0.02469240128993988, + 0.010204299353063107, + 0.005760228261351585, + 0.008338247425854206, + 0.003580261953175068, + 0.010471154004335403, + 0.004760500509291887, + 0.0064245229586958885, + 0.00479878531768918, + 0.006769402418285608, + 0.006255917716771364, + 0.005660128779709339, + 0.003888338804244995, + 0.007143849972635508, + 0.006703848019242287, + 0.03517095744609833, + 0.04962673410773277, + 0.04546092823147774, + 0.07328888773918152, + 0.04226407781243324, + 0.02846573293209076, + 0.05185304954648018, + 0.035262223333120346, + 0.010839532129466534, + 0.011686724610626698, + 0.009031863883137703, + 0.0064796037040650845, + 0.006693732459098101, + 0.010158638469874859, + 0.01141513790935278, + 0.011345586739480495, + 0.004726490005850792, + 0.007542577106505632, + 0.00470745051279664, + 0.005202505271881819, + 0.004799933638423681, + 0.006641690619289875, + 0.005539919249713421, + 0.006930885370820761, + 0.014813445508480072, + 0.02529832534492016, + 0.016908278688788414, + 0.03358139842748642, + 0.029259705916047096, + 0.017573941498994827, + 0.022690778598189354, + 0.01967611350119114, + 0.0017466115532442927, + 0.0012405022280290723, + 0.0032013445161283016, + 0.001830088789574802, + 0.001853223075158894, + 0.0015026632463559508, + 0.01293340977281332, + 0.011770281009376049, + 0.004084781277924776, + 0.0022670412436127663, + 0.0056758145801723, + 0.008147899992763996, + 0.0039843306876719, + 0.0034595662727952003, + 0.017862027511000633, + 0.013017470017075539, + 0.006262991111725569, + 0.003561393590644002, + 0.0059629883617162704, + 0.008096138946712017, + 0.015899330377578735, + 0.007350672502070665, + 0.008910423144698143, + 0.010441037826240063, + 0.0032840503845363855, + 0.0033272646833211184, + 0.004263722337782383, + 0.005904915276914835, + 0.0043963538482785225, + 0.0033986230846494436, + 0.005029483698308468, + 0.003793153213337064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.023281434550881386 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, handheld, medium shot panning rightward plunges the viewer into the depths of a deep, dark cave. At the center of the frame is an explorer with tousled dark brown hair and a face stained with dirt. His expression is focused and alert. He wears a frayed dark green canvas jacket, layered over a gray Henley shirt, its surface darkened by the cave's dampness. In his left hand, he holds a burning torch. The dancing flame, the only source of light in the pitch black, crackles continuously and casts a warm, dynamic orange-yellow glow on the surroundings. This unstable light casts swaying, elongated shadows on the damp, craggy rock face, covered in shimmering moss and dripping with crystal-clear water droplets. He moves cautiously along the rock face, leaning slightly forward, his right hand extended, his fingertips gently and cautiously touching the slippery surface. His gaze follows the movement of his finger, scanning for any crack or unusual protrusion that might hold a clue. The entire scene creates an atmosphere of tension, mystery, and exploration with dynamic lighting, focused movements, and the crackling of torches.", + "seed": 7054, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_10/index_9484_seed_7054_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09898190733511, + "vlm_logprob": -0.09898190733511, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.09898190733511, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.36923491954803467 + ], + "opposite_logprob_tokens": [ + -1.2442349195480347 + ], + "logprob_reward": -0.36923491954803467, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.36923491954803467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3615243434906006, + 0.20558765530586243, + 0.05842326581478119, + 0.028181105852127075, + 0.05352342128753662, + 0.06927001476287842, + 0.07209499925374985, + 0.06246712803840637, + 0.06187870353460312, + 0.02704937569797039 + ], + "credit_map": [ + 0.06748127937316895, + 0.008829515427350998, + 0.03494207188487053, + 0.018821390345692635, + 0.028412509709596634, + 0.03725380077958107, + 0.14342845976352692, + 0.022355325520038605, + 0.030535424128174782, + 0.0046691554598510265, + 0.08118817955255508, + 0.01238317135721445, + 0.008893435820937157, + 0.00966212060302496, + 0.047067269682884216, + 0.011188899166882038, + 0.004508358892053366, + 0.006249471567571163, + 0.008998658508062363, + 0.00884419959038496, + 0.00735735148191452, + 0.009435605257749557, + 0.006863993592560291, + 0.006165630649775267, + 0.0019654487259685993, + 0.0030628731474280357, + 0.004259740933775902, + 0.006260788533836603, + 0.0025339757557958364, + 0.002569032134488225, + 0.004227721132338047, + 0.0033015268854796886, + 0.0036944488529115915, + 0.005181894637644291, + 0.007408480159938335, + 0.008794751949608326, + 0.007091723382472992, + 0.006751961540430784, + 0.006997730117291212, + 0.00760243134573102, + 0.00593680702149868, + 0.00820126198232174, + 0.007421750109642744, + 0.008372860960662365, + 0.006191132124513388, + 0.011583850719034672, + 0.01203575637191534, + 0.009526592679321766, + 0.006513126660138369, + 0.005267574917525053, + 0.008349488489329815, + 0.013228715397417545, + 0.00842976476997137, + 0.010801383294165134, + 0.012544592842459679, + 0.006960354745388031, + 0.00641645397990942, + 0.003633589018136263, + 0.007237417623400688, + 0.009954935871064663, + 0.005965016782283783, + 0.01254984736442566, + 0.010114642791450024, + 0.006595222279429436, + 0.005672600585967302, + 0.007119812071323395, + 0.006103524938225746, + 0.00836874172091484, + 0.009370830841362476, + 0.014059609733521938, + 0.005399088840931654, + 0.005784497130662203, + 0.002321346662938595, + 0.0017901038518175483, + 0.002489538164809346, + 0.003731005359441042, + 0.004007683135569096, + 0.0042159841395914555, + 0.005511221941560507, + 0.0029824930243194103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.36923491954803467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048715412616729736 + ], + "opposite_logprob_tokens": [ + -3.048715353012085 + ], + "logprob_reward": -0.048715412616729736, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048715412616729736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3170550465583801, + 0.13229000568389893, + 0.051409948617219925, + 0.04708198085427284, + 0.11304307729005814, + 0.07348962128162384, + 0.05734102055430412, + 0.11273492127656937, + 0.04195067286491394, + 0.053603775799274445 + ], + "credit_map": [ + 0.059398800134658813, + 0.013932056725025177, + 0.029983170330524445, + 0.029672371223568916, + 0.0292037483304739, + 0.025737514719367027, + 0.1014876738190651, + 0.0276397205889225, + 0.017015371471643448, + 0.008826342411339283, + 0.015145857818424702, + 0.01600736938416958, + 0.014182869344949722, + 0.020627042278647423, + 0.01944972388446331, + 0.021035421639680862, + 0.004623780958354473, + 0.005696136504411697, + 0.0069462088868021965, + 0.006153659429401159, + 0.005506431218236685, + 0.010806389153003693, + 0.006048096343874931, + 0.005629241932183504, + 0.005230891052633524, + 0.0046443152241408825, + 0.005190607625991106, + 0.0070152804255485535, + 0.0063205985352396965, + 0.005805172026157379, + 0.008126242086291313, + 0.004748872481286526, + 0.007666357327252626, + 0.013080704025924206, + 0.014724415726959705, + 0.01958213374018669, + 0.011022915132343769, + 0.013368860818445683, + 0.01570853218436241, + 0.017889153212308884, + 0.006579049397259951, + 0.005072633735835552, + 0.007418786641210318, + 0.00958766508847475, + 0.006551251746714115, + 0.00995105504989624, + 0.019293084740638733, + 0.009036090224981308, + 0.005946275778114796, + 0.006387011148035526, + 0.00954680610448122, + 0.010646511800587177, + 0.004766453988850117, + 0.008694096468389034, + 0.006206548772752285, + 0.005147316958755255, + 0.01072682999074459, + 0.009215637110173702, + 0.014575409702956676, + 0.023827893659472466, + 0.012243532575666904, + 0.0226078350096941, + 0.009558304212987423, + 0.009979491122066975, + 0.003376880893483758, + 0.005225429777055979, + 0.00532932672649622, + 0.00492460560053587, + 0.00652087340131402, + 0.007117752451449633, + 0.004663374740630388, + 0.004792430438101292, + 0.004953161347657442, + 0.0036889645271003246, + 0.006980185396969318, + 0.0057219890877604485, + 0.007700788322836161, + 0.005564989056438208, + 0.013140209019184113, + 0.00585348904132843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.048715412616729736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026364656165242195 + ], + "opposite_logprob_tokens": [ + -3.65136456489563 + ], + "logprob_reward": -0.026364656165242195, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026364656165242195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.548945963382721, + 0.08337447792291641, + 0.04306866601109505, + 0.06636509299278259, + 0.03544170781970024, + 0.05568134784698486, + 0.04454736039042473, + 0.03149324283003807, + 0.032282259315252304, + 0.05879979953169823 + ], + "credit_map": [ + 0.11040256172418594, + 0.012560958042740822, + 0.05048039183020592, + 0.024232186377048492, + 0.04447155445814133, + 0.054176315665245056, + 0.21696554124355316, + 0.03565649315714836, + 0.01268515083938837, + 0.008101530373096466, + 0.02345888316631317, + 0.008270075544714928, + 0.007801462896168232, + 0.006486255209892988, + 0.011259003542363644, + 0.005312127061188221, + 0.0035007691476494074, + 0.005070304498076439, + 0.005684174597263336, + 0.006358474493026733, + 0.006145494524389505, + 0.006788755767047405, + 0.004885345697402954, + 0.004635347984731197, + 0.006415464449673891, + 0.005657871253788471, + 0.005952755454927683, + 0.008140266872942448, + 0.011388781480491161, + 0.013559487648308277, + 0.008969384245574474, + 0.006281091831624508, + 0.0029682174790650606, + 0.0032614117953926325, + 0.003481250489130616, + 0.0032838499173521996, + 0.00599708640947938, + 0.006642836146056652, + 0.004930625669658184, + 0.004876433871686459, + 0.006173443049192429, + 0.0063024284318089485, + 0.004893636330962181, + 0.007556220982223749, + 0.004968138877302408, + 0.006492780987173319, + 0.010468045249581337, + 0.008826655335724354, + 0.0037861664313822985, + 0.003610603278502822, + 0.006472608540207148, + 0.007433472201228142, + 0.004536770284175873, + 0.00767940329387784, + 0.006925519555807114, + 0.004102821461856365, + 0.0027442106511443853, + 0.0025170401204377413, + 0.004394978284835815, + 0.005793116521090269, + 0.003551033092662692, + 0.006550473161041737, + 0.0026370747946202755, + 0.00330531457439065, + 0.0029753725975751877, + 0.003288906067609787, + 0.004130491986870766, + 0.00339787220582366, + 0.00565143721178174, + 0.004452207125723362, + 0.005128675140440464, + 0.003257298842072487, + 0.006375494413077831, + 0.006170215085148811, + 0.006009337492287159, + 0.00663808872923255, + 0.009087336249649525, + 0.009109808132052422, + 0.009455318562686443, + 0.005954204127192497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.026364656165242195 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019417019793763757 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019417019793763757, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019417019793763757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5614325404167175, + 0.17513129115104675, + 0.020305689424276352, + 0.026400752365589142, + 0.070456862449646, + 0.04114863649010658, + 0.029343029484152794, + 0.03768421709537506, + 0.023875167593359947, + 0.014221776276826859 + ], + "credit_map": [ + 0.11393976956605911, + 0.0127534419298172, + 0.05289347097277641, + 0.02574647217988968, + 0.04442744329571724, + 0.05104351416230202, + 0.22609753906726837, + 0.03453090041875839, + 0.026021890342235565, + 0.004430508706718683, + 0.06979706138372421, + 0.011437937617301941, + 0.007437066175043583, + 0.006925545632839203, + 0.040908828377723694, + 0.00817243941128254, + 0.001970636425539851, + 0.0016246046870946884, + 0.002278246683999896, + 0.0020415163598954678, + 0.002215693239122629, + 0.005220445338636637, + 0.0026458746287971735, + 0.002308671362698078, + 0.001601488096639514, + 0.002017286606132984, + 0.0029150829650461674, + 0.0029446587432175875, + 0.004836746957153082, + 0.004281254019588232, + 0.004205027595162392, + 0.003599204821512103, + 0.004185685887932777, + 0.004560955800116062, + 0.006537444889545441, + 0.008370212279260159, + 0.01567910611629486, + 0.011650369502604008, + 0.009537163190543652, + 0.009935923852026463, + 0.003653940511867404, + 0.0027926464099437, + 0.005110211670398712, + 0.005010481458157301, + 0.006421241909265518, + 0.005406551994383335, + 0.008244358003139496, + 0.004509203601628542, + 0.0035532342735677958, + 0.002287533599883318, + 0.0040574041195213795, + 0.005973861087113619, + 0.00252862018533051, + 0.004394779913127422, + 0.003812043694779277, + 0.0027355547063052654, + 0.0033787910360842943, + 0.0026742478366941214, + 0.0034902351908385754, + 0.005193078890442848, + 0.004851542878895998, + 0.005454214755445719, + 0.007889344356954098, + 0.004752761218696833, + 0.00221757753752172, + 0.0020923707634210587, + 0.003889970015734434, + 0.002936950186267495, + 0.0033264714293181896, + 0.003114317776635289, + 0.0026020361110568047, + 0.0036954728420823812, + 0.001136727980338037, + 0.0010339668951928616, + 0.0014003003016114235, + 0.0015419855481013656, + 0.0027345893904566765, + 0.0031753117218613625, + 0.0018859414849430323, + 0.0013129522558301687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0019417019793763757 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that pans rightward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the explorer wearing a frayed dark green canvas jacket over a gray Henley shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the explorer holding a burning torch in his left hand that provides the only light source?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a deep, dark cave with damp, craggy rock faces covered in moss and water droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the explorer moving cautiously while touching the rock face with his right fingertips?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865284636616707 + ], + "opposite_logprob_tokens": [ + -3.0486528873443604 + ], + "logprob_reward": -0.04865284636616707, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865284636616707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5142336487770081, + 0.0986001119017601, + 0.024406973272562027, + 0.02526303008198738, + 0.03767693415284157, + 0.19421136379241943, + 0.02208607830107212, + 0.038255415856838226, + 0.027297770604491234, + 0.017968710511922836 + ], + "credit_map": [ + 0.10096228867769241, + 0.012112665921449661, + 0.04771092161536217, + 0.02328435704112053, + 0.04197853058576584, + 0.05145592987537384, + 0.2056235373020172, + 0.03110540844500065, + 0.015922732651233673, + 0.0036404721904546022, + 0.03843814507126808, + 0.0067259580828249454, + 0.004293221514672041, + 0.0036977052222937346, + 0.02068827673792839, + 0.005193591117858887, + 0.0024014082737267017, + 0.00261169602163136, + 0.003541408572345972, + 0.004098350647836924, + 0.002739380579441786, + 0.0034754639491438866, + 0.002988820429891348, + 0.0025504429358989, + 0.001988210715353489, + 0.0025528904516249895, + 0.00416948739439249, + 0.004428992047905922, + 0.002673555165529251, + 0.0025277328677475452, + 0.0038014831952750683, + 0.0031206775456666946, + 0.004511361941695213, + 0.005748223979026079, + 0.0038029623683542013, + 0.005631359294056892, + 0.003916559275239706, + 0.005123829934746027, + 0.004123917780816555, + 0.0048187184147536755, + 0.017279449850320816, + 0.012865317054092884, + 0.016194652765989304, + 0.03257704898715019, + 0.014960432425141335, + 0.03278040513396263, + 0.0460725836455822, + 0.021481459960341454, + 0.0017767902463674545, + 0.0011870991438627243, + 0.002387100365012884, + 0.0027733901515603065, + 0.003448690753430128, + 0.003632122650742531, + 0.0045682587660849094, + 0.0023126250598579645, + 0.004312385339289904, + 0.003128827316686511, + 0.004996747709810734, + 0.007486043497920036, + 0.003768750699236989, + 0.0070711406879127026, + 0.00424625538289547, + 0.0032452652230858803, + 0.002119575161486864, + 0.004006421659141779, + 0.002757397247478366, + 0.004552057012915611, + 0.0036685869563370943, + 0.005075901746749878, + 0.0022901177871972322, + 0.0028277121018618345, + 0.0021237649489194155, + 0.0019132280722260475, + 0.0019237479427829385, + 0.0025758082047104836, + 0.0021219314076006413, + 0.0024036169052124023, + 0.002501504495739937, + 0.002405107021331787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.04865284636616707 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_11/index_tensor([5572, 5572])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_11/index_tensor([5572, 5572])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..86715f8b3220fd11cfe62f42b330e511bfaacbf4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_11/index_tensor([5572, 5572])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level extreme close-up captures a classic, clear pint glass placed directly on the smooth, dark mahogany surface of a bar. The glass is filled with rich, amber-colored malt liquor, and a thick, creamy, off-white foam pours relentlessly, rising to the rim and beginning its slow, silent descent. The foam cascades down the curved sides of the glass in a thick stream of bubbles, moving at a steady, deliberate pace until it gathers in an ever-expanding, lace-like pool on the polished wood below. The entire scene is dramatically illuminated from outside the frame by a warm, low-hanging light source, casting a focused glow that accentuates the glistening, damp texture of the foam and creates bright, sharp reflections on the bar's surface, while plunging the softly blurred background into deep shadow. The atmosphere is one of quiet, lonely neglect, a wasted yet mesmerizingly beautiful moment captured in a tightly framed, intimate shot.", + "seed": 28433, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_11/index_5572_seed_28433_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0200598802883178, + "vlm_logprob": -0.0200598802883178, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.0200598802883178, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010059256106615067 + ], + "opposite_logprob_tokens": [ + -5.135059356689453 + ], + "logprob_reward": -0.010059256106615067, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010059256106615067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17313049733638763, + 0.09527234733104706, + 0.05837137624621391, + 0.03215420991182327, + 0.06926201283931732, + 0.06454796344041824, + 0.3135164976119995, + 0.054319705814123154, + 0.02641606144607067, + 0.1130092591047287 + ], + "credit_map": [ + 0.020158283412456512, + 0.019754763692617416, + 0.030947184190154076, + 0.02529488317668438, + 0.014818022027611732, + 0.01596703752875328, + 0.023210003972053528, + 0.022980330511927605, + 0.013001077808439732, + 0.008961710147559643, + 0.015937810763716698, + 0.016094958409667015, + 0.006372644565999508, + 0.011355800554156303, + 0.01202382892370224, + 0.011524523608386517, + 0.009852518327534199, + 0.005545589607208967, + 0.007682764437049627, + 0.010149198584258556, + 0.005504647269845009, + 0.00723111629486084, + 0.007330010645091534, + 0.005075535736978054, + 0.005951321683824062, + 0.0029251067899167538, + 0.00543568329885602, + 0.003999270964413881, + 0.0028529928531497717, + 0.004421541001647711, + 0.0030772550962865353, + 0.003491041250526905, + 0.01016337238252163, + 0.00511765293776989, + 0.010809633880853653, + 0.009514596313238144, + 0.005804344546049833, + 0.008118157275021076, + 0.0117794806137681, + 0.007954775355756283, + 0.007480908650904894, + 0.0035921186208724976, + 0.01229867897927761, + 0.007311870809644461, + 0.005683524068444967, + 0.007786087226122618, + 0.014845917001366615, + 0.005548865534365177, + 0.042812321335077286, + 0.018286824226379395, + 0.06541259586811066, + 0.036749813705682755, + 0.021531911566853523, + 0.04070243611931801, + 0.061446502804756165, + 0.02657410502433777, + 0.007600790821015835, + 0.00563505943864584, + 0.00525754876434803, + 0.004764375742524862, + 0.007424044888466597, + 0.008829804137349129, + 0.004871144890785217, + 0.009936941787600517, + 0.003200413892045617, + 0.0036849789321422577, + 0.004845513496547937, + 0.002715726848691702, + 0.0030728208366781473, + 0.003394092433154583, + 0.003129290184006095, + 0.0023732283152639866, + 0.018605010583996773, + 0.009406437166035175, + 0.017819862812757492, + 0.008589942939579487, + 0.020097309723496437, + 0.01949402317404747, + 0.009435865096747875, + 0.009560824371874332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.010059256106615067 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608537696301937 + ], + "opposite_logprob_tokens": [ + -4.141085147857666 + ], + "logprob_reward": -0.01608537696301937, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608537696301937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2281249314546585, + 0.1298605054616928, + 0.1172381341457367, + 0.020479975268244743, + 0.048879027366638184, + 0.027928493916988373, + 0.11199156939983368, + 0.1646251678466797, + 0.031661368906497955, + 0.11921083182096481 + ], + "credit_map": [ + 0.035521604120731354, + 0.020471537485718727, + 0.04372812807559967, + 0.028557904064655304, + 0.023580266162753105, + 0.026183292269706726, + 0.026026548817753792, + 0.02405565232038498, + 0.021792929619550705, + 0.006020268891006708, + 0.04328205808997154, + 0.011361614800989628, + 0.008590955287218094, + 0.020973917096853256, + 0.006602626759558916, + 0.01123612280935049, + 0.015282915905117989, + 0.009011784568428993, + 0.015393711626529694, + 0.020748727023601532, + 0.00917858351022005, + 0.01382859330624342, + 0.017832061275839806, + 0.015961745753884315, + 0.0034634554758667946, + 0.00178700836841017, + 0.002885515335947275, + 0.002391195623204112, + 0.0018514953553676605, + 0.0026324274949729443, + 0.002553313970565796, + 0.0029155639931559563, + 0.007286753971129656, + 0.003515565535053611, + 0.008314266800880432, + 0.007128198631107807, + 0.0036324714310467243, + 0.005903723184019327, + 0.0071791065856814384, + 0.005918940529227257, + 0.004074920900166035, + 0.002221105620265007, + 0.006239117588847876, + 0.0033378812950104475, + 0.0023852274753153324, + 0.0034443270415067673, + 0.0033691436983644962, + 0.0028567679692059755, + 0.017396539449691772, + 0.006453732959926128, + 0.0287147406488657, + 0.01131900679320097, + 0.007927975617349148, + 0.01666686125099659, + 0.014697737991809845, + 0.008814962580800056, + 0.018036358058452606, + 0.021307116374373436, + 0.027105141431093216, + 0.01429678499698639, + 0.02561165764927864, + 0.019812174141407013, + 0.012442030012607574, + 0.026013894006609917, + 0.0050811986438930035, + 0.003286721184849739, + 0.0034102718345820904, + 0.00327734905295074, + 0.0040986062958836555, + 0.0053569781593978405, + 0.0031889937818050385, + 0.003961246460676193, + 0.01850014552474022, + 0.011544606648385525, + 0.014073074795305729, + 0.007903585210442543, + 0.022228848189115524, + 0.01967160403728485, + 0.012919764965772629, + 0.012369186617434025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.01608537696301937 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048666246235370636 + ], + "opposite_logprob_tokens": [ + -3.04866623878479 + ], + "logprob_reward": -0.048666246235370636, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048666246235370636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24836686253547668, + 0.13483160734176636, + 0.13082006573677063, + 0.04280034452676773, + 0.07356442511081696, + 0.04085303097963333, + 0.058598779141902924, + 0.0757870152592659, + 0.059065066277980804, + 0.1353127658367157 + ], + "credit_map": [ + 0.040014464408159256, + 0.019657673314213753, + 0.04633157700300217, + 0.034422218799591064, + 0.015668485313653946, + 0.028961626812815666, + 0.033228859305381775, + 0.030081968754529953, + 0.023247675970196724, + 0.005409849341958761, + 0.04642263054847717, + 0.01082552969455719, + 0.008080589585006237, + 0.02150758169591427, + 0.008369878865778446, + 0.010967879556119442, + 0.017843522131443024, + 0.00947517342865467, + 0.021939221769571304, + 0.01887042634189129, + 0.008158084005117416, + 0.01743890531361103, + 0.016917651519179344, + 0.020177079364657402, + 0.007613949943333864, + 0.004146011080592871, + 0.006423301063477993, + 0.006169036030769348, + 0.004205276258289814, + 0.005463872570544481, + 0.0044906483963131905, + 0.004288249183446169, + 0.013061704114079475, + 0.004059316590428352, + 0.01820351369678974, + 0.008422546088695526, + 0.0050253053195774555, + 0.010755590163171291, + 0.00783843919634819, + 0.00619801040738821, + 0.00571183068677783, + 0.0036557551939040422, + 0.009574766270816326, + 0.005051464308053255, + 0.003438264364376664, + 0.004697385244071484, + 0.005033115390688181, + 0.0036904492881149054, + 0.00984807126224041, + 0.004362504929304123, + 0.012762918137013912, + 0.004727897699922323, + 0.005785477347671986, + 0.008829953148961067, + 0.005007302388548851, + 0.007274657487869263, + 0.00846811942756176, + 0.012036032974720001, + 0.009140752255916595, + 0.0054671745747327805, + 0.0150935472920537, + 0.010918410494923592, + 0.0077031925320625305, + 0.0069597880356013775, + 0.007499681320041418, + 0.006756380666047335, + 0.007676885463297367, + 0.0048351772129535675, + 0.008377447724342346, + 0.008091357536613941, + 0.0067359209060668945, + 0.009092213585972786, + 0.02334476262331009, + 0.011417459696531296, + 0.013514462858438492, + 0.0076591805554926395, + 0.026707151904702187, + 0.02552798204123974, + 0.013051102869212627, + 0.014090664684772491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.048666246235370636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287957534193993 + ], + "opposite_logprob_tokens": [ + -3.7732880115509033 + ], + "logprob_reward": -0.023287957534193993, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287957534193993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17854976654052734, + 0.14246754348278046, + 0.13509953022003174, + 0.046373434364795685, + 0.06958511471748352, + 0.1515667885541916, + 0.08667471259832382, + 0.0923394188284874, + 0.019873812794685364, + 0.07746991515159607 + ], + "credit_map": [ + 0.022612109780311584, + 0.01844320446252823, + 0.03387745842337608, + 0.02683090791106224, + 0.011530948802828789, + 0.01677797921001911, + 0.021814929321408272, + 0.026662223041057587, + 0.020085755735635757, + 0.009225573390722275, + 0.03517312556505203, + 0.020090462639927864, + 0.0084599070250988, + 0.02080298587679863, + 0.012366308830678463, + 0.01626342535018921, + 0.021078214049339294, + 0.0105708297342062, + 0.023358000442385674, + 0.016254259273409843, + 0.009800640866160393, + 0.017883669584989548, + 0.013454619795084, + 0.022699294611811638, + 0.005749473348259926, + 0.004017076920717955, + 0.008146805688738823, + 0.00779162161052227, + 0.003529686015099287, + 0.005148120690137148, + 0.005105852149426937, + 0.00688480120152235, + 0.010004991665482521, + 0.006354302633553743, + 0.010364464484155178, + 0.008626259863376617, + 0.005870104301720858, + 0.008184035308659077, + 0.010636148042976856, + 0.00954480655491352, + 0.022226789966225624, + 0.007465629372745752, + 0.044174063950777054, + 0.016538569703698158, + 0.010134104639291763, + 0.021535998210310936, + 0.017062416300177574, + 0.012429224327206612, + 0.011214790865778923, + 0.005522537510842085, + 0.013722063042223454, + 0.011553647927939892, + 0.006311706732958555, + 0.009634609334170818, + 0.018299907445907593, + 0.010415456257760525, + 0.01172614935785532, + 0.01156295370310545, + 0.011279030703008175, + 0.0101926876232028, + 0.012224842794239521, + 0.012107360176742077, + 0.008839059621095657, + 0.014407332986593246, + 0.003316132817417383, + 0.0015175237786024809, + 0.004558585584163666, + 0.0018654931336641312, + 0.0018308490980416536, + 0.00307549093849957, + 0.0017179122660309076, + 0.0019918233156204224, + 0.011568645015358925, + 0.006654357071965933, + 0.015151538886129856, + 0.006595986429601908, + 0.012411392293870449, + 0.011002068407833576, + 0.006563697475939989, + 0.007522231433540583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.023287957534193993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022005646023899317 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.0022005646023899317, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022005646023899317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16640865802764893, + 0.14924412965774536, + 0.18155156075954437, + 0.05300254374742508, + 0.09196698665618896, + 0.15598449110984802, + 0.0494937002658844, + 0.05275953188538551, + 0.03697674348950386, + 0.06261162459850311 + ], + "credit_map": [ + 0.031126830726861954, + 0.011534687131643295, + 0.03344116732478142, + 0.021002983674407005, + 0.010673566721379757, + 0.02250787802040577, + 0.01734994165599346, + 0.01877160556614399, + 0.024639008566737175, + 0.007222102023661137, + 0.040362704545259476, + 0.017110086977481842, + 0.008961288258433342, + 0.025177570059895515, + 0.010627811774611473, + 0.015143567696213722, + 0.028091199696063995, + 0.01116945967078209, + 0.03390050679445267, + 0.022037241607904434, + 0.012697908096015453, + 0.022648833692073822, + 0.020975222811102867, + 0.03003118745982647, + 0.007832290604710579, + 0.0036076840478926897, + 0.009697433561086655, + 0.0080022644251585, + 0.0044417413882911205, + 0.006282306741923094, + 0.005811659619212151, + 0.007327161263674498, + 0.014231699518859386, + 0.010563966818153858, + 0.01719544641673565, + 0.007941189222037792, + 0.008760177530348301, + 0.012084543704986572, + 0.01193968579173088, + 0.009250272996723652, + 0.02105739712715149, + 0.009282388724386692, + 0.050337888300418854, + 0.0180004071444273, + 0.00927133671939373, + 0.0231234822422266, + 0.010658410377800465, + 0.014253181405365467, + 0.00823422521352768, + 0.0045128148049116135, + 0.005009832326322794, + 0.005939811933785677, + 0.005147566087543964, + 0.007372959516942501, + 0.008095895871520042, + 0.0051805973052978516, + 0.007452495861798525, + 0.005100579932332039, + 0.010145491920411587, + 0.00533602200448513, + 0.0042180586606264114, + 0.007233065087348223, + 0.004430872388184071, + 0.008842943236231804, + 0.004766559693962336, + 0.004788424354046583, + 0.003853687783703208, + 0.002782321535050869, + 0.004826315678656101, + 0.0061948844231665134, + 0.004493813030421734, + 0.005270733498036861, + 0.010734757408499718, + 0.006121509708464146, + 0.00690554128959775, + 0.004873873200267553, + 0.012529037892818451, + 0.010433646850287914, + 0.005383041687309742, + 0.005630220286548138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0022005646023899317 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level extreme close-up captures a classic, clear pint glass placed directly on the smooth, dark mahogany surface of a bar. The glass is filled with rich, amber-colored malt liquor, and a thick, creamy, off-white foam pours relentlessly, rising to the rim and beginning its slow, silent descent. The foam cascades down the curved sides of the glass in a thick stream of bubbles, moving at a steady, deliberate pace until it gathers in an ever-expanding, lace-like pool on the polished wood below. The entire scene is dramatically illuminated from outside the frame by a warm, low-hanging light source, casting a focused glow that accentuates the glistening, damp texture of the foam and creates bright, sharp reflections on the bar's surface, while plunging the softly blurred background into deep shadow. The atmosphere is one of quiet, lonely neglect, a wasted yet mesmerizingly beautiful moment captured in a tightly framed, intimate shot.", + "seed": 28434, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_11/index_5572_seed_28434_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012362546613439917, + "vlm_logprob": -0.012362546613439917, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.012362546613439917, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00783595908433199 + ], + "opposite_logprob_tokens": [ + -5.382835865020752 + ], + "logprob_reward": -0.00783595908433199, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00783595908433199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13728365302085876, + 0.05826502665877342, + 0.03760555386543274, + 0.06995249539613724, + 0.1955689638853073, + 0.03091742843389511, + 0.2951491177082062, + 0.07169017940759659, + 0.08061110228300095, + 0.022956475615501404 + ], + "credit_map": [ + 0.03992001712322235, + 0.005012099165469408, + 0.018477212637662888, + 0.00438873004168272, + 0.00759486947208643, + 0.014466820284724236, + 0.007329949643462896, + 0.040093954652547836, + 0.009113632142543793, + 0.005103656556457281, + 0.009729241952300072, + 0.0075842575170099735, + 0.005354638211429119, + 0.007369993720203638, + 0.006584153976291418, + 0.007425450719892979, + 0.007140937261283398, + 0.002389490371569991, + 0.004361436702311039, + 0.0032350108958780766, + 0.005110267549753189, + 0.003700210712850094, + 0.004224228207021952, + 0.007443971466273069, + 0.01832060143351555, + 0.0041419388726353645, + 0.007912378758192062, + 0.003853759029880166, + 0.005089630838483572, + 0.007232060190290213, + 0.004443954676389694, + 0.01895817182958126, + 0.06098080798983574, + 0.005995100364089012, + 0.012199226766824722, + 0.004443417768925428, + 0.004758016671985388, + 0.01778745837509632, + 0.010201001539826393, + 0.07920393347740173, + 0.007246760185807943, + 0.001974196406081319, + 0.0041168383322656155, + 0.0018770734313875437, + 0.0030114687979221344, + 0.0032115813810378313, + 0.0028334055095911026, + 0.00664610555395484, + 0.11300049722194672, + 0.008095160126686096, + 0.011300794780254364, + 0.006897856015712023, + 0.01073444727808237, + 0.12480194121599197, + 0.009067729115486145, + 0.01125070545822382, + 0.023839756846427917, + 0.0030239929910749197, + 0.0034247792791575193, + 0.0022568462882190943, + 0.0038082879036664963, + 0.025684112682938576, + 0.004896142520010471, + 0.004756262991577387, + 0.031124627217650414, + 0.002401459962129593, + 0.0025150664150714874, + 0.0015760068781673908, + 0.002515627769753337, + 0.03401126712560654, + 0.00318288360722363, + 0.0032841654028743505, + 0.00805404782295227, + 0.0008691325783729553, + 0.0009290172602050006, + 0.000838019885122776, + 0.000930194219108671, + 0.008879687637090683, + 0.0012957672588527203, + 0.0011606091866269708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.00783595908433199 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004648711532354355 + ], + "opposite_logprob_tokens": [ + -5.379648685455322 + ], + "logprob_reward": -0.004648711532354355, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004648711532354355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040211912244558334, + 0.016181889921426773, + 0.07457069307565689, + 0.08394945412874222, + 0.12653888761997223, + 0.030947567895054817, + 0.07831104099750519, + 0.10207759588956833, + 0.05175115913152695, + 0.3954598009586334 + ], + "credit_map": [ + 0.006997678428888321, + 0.0039873551577329636, + 0.0069953808560967445, + 0.002853004727512598, + 0.005467651411890984, + 0.00357934576459229, + 0.004987119231373072, + 0.0053443750366568565, + 0.002024272456765175, + 0.0021115096751600504, + 0.0023963351268321276, + 0.0020846312399953604, + 0.001952815568074584, + 0.0013029398396611214, + 0.0021257016342133284, + 0.0021836822852492332, + 0.0144934868440032, + 0.004768856335431337, + 0.006080042105168104, + 0.006817826069891453, + 0.008494244888424873, + 0.0077849989756941795, + 0.012986637651920319, + 0.013144594617187977, + 0.018934082239866257, + 0.0065222084522247314, + 0.007575882598757744, + 0.0056008887477219105, + 0.00937821064144373, + 0.009269545786082745, + 0.008402186445891857, + 0.018266450613737106, + 0.039214227348566055, + 0.003636871464550495, + 0.007777890656143427, + 0.0026417935732752085, + 0.0029953117482364178, + 0.011826595291495323, + 0.006107456050813198, + 0.052338749170303345, + 0.008024572394788265, + 0.0017567428294569254, + 0.002636018441990018, + 0.0019426675280556083, + 0.0022433195263147354, + 0.005175817757844925, + 0.003791877068579197, + 0.005376552231609821, + 0.027626199647784233, + 0.002658869605511427, + 0.003086038399487734, + 0.002269779797643423, + 0.005663616582751274, + 0.030624819919466972, + 0.003075944958254695, + 0.0033057767432183027, + 0.03796732425689697, + 0.002927718684077263, + 0.003889132756739855, + 0.002222178503870964, + 0.0034510106779634953, + 0.04168885573744774, + 0.004775515291839838, + 0.005155861843377352, + 0.020036770030856133, + 0.0015151264378800988, + 0.0017685126513242722, + 0.000995395239442587, + 0.0016097776824608445, + 0.022022195160388947, + 0.0017264471389353275, + 0.002076931996271014, + 0.15908290445804596, + 0.009601477533578873, + 0.010952326469123363, + 0.005760896950960159, + 0.010204005055129528, + 0.1753089427947998, + 0.010512114502489567, + 0.014037154614925385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.004648711532354355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187422305345535 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187422305345535, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187422305345535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06134408339858055, + 0.04005924239754677, + 0.02324562892317772, + 0.05329132452607155, + 0.06520894914865494, + 0.02848002128303051, + 0.048803020268678665, + 0.1368885487318039, + 0.07400801032781601, + 0.46867111325263977 + ], + "credit_map": [ + 0.011273548938333988, + 0.006368005648255348, + 0.008943408727645874, + 0.004238844383507967, + 0.008999636396765709, + 0.007156203966587782, + 0.006751082371920347, + 0.00761335389688611, + 0.005952499806880951, + 0.0033524883911013603, + 0.005814017727971077, + 0.0040171486325562, + 0.005707071628421545, + 0.004305921960622072, + 0.005754981189966202, + 0.005155112128704786, + 0.004881064407527447, + 0.0019249825272709131, + 0.0016232654452323914, + 0.0015357705997303128, + 0.0035112623590976, + 0.0023421398364007473, + 0.0028863579500466585, + 0.0045407856814563274, + 0.011468681506812572, + 0.003897062037140131, + 0.0053843422792851925, + 0.0029415327589958906, + 0.008703773841261864, + 0.006429980508983135, + 0.005152728408575058, + 0.009313222020864487, + 0.01797405630350113, + 0.002715593669563532, + 0.004335793666541576, + 0.0016065408708527684, + 0.0028928834944963455, + 0.006381507031619549, + 0.003914655186235905, + 0.025387918576598167, + 0.005209286697208881, + 0.0017545795999467373, + 0.002635343000292778, + 0.0026787505485117435, + 0.0016429416136816144, + 0.005839740391820669, + 0.005173887126147747, + 0.0035454914905130863, + 0.017115645110607147, + 0.0014883666299283504, + 0.0028643575496971607, + 0.0012419522972777486, + 0.002326281275600195, + 0.018473204225301743, + 0.002451026812195778, + 0.002842181595042348, + 0.05286693200469017, + 0.003623181488364935, + 0.004394222982227802, + 0.002579821739345789, + 0.005007445812225342, + 0.05778329074382782, + 0.005224841646850109, + 0.005408809054642916, + 0.027422737330198288, + 0.001981243724003434, + 0.002287261188030243, + 0.0016317712143063545, + 0.0031031703110784292, + 0.029940485954284668, + 0.004217248875647783, + 0.0034240952227264643, + 0.18868593871593475, + 0.011415495537221432, + 0.013063961640000343, + 0.006851385813206434, + 0.01209036260843277, + 0.2078474462032318, + 0.01233119796961546, + 0.016385339200496674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.018187422305345535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803529381752014 + ], + "opposite_logprob_tokens": [ + -3.529803514480591 + ], + "logprob_reward": -0.029803529381752014, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803529381752014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1325213760137558, + 0.026613736525177956, + 0.01660078950226307, + 0.03807118535041809, + 0.08387180417776108, + 0.024062631651759148, + 0.09909021854400635, + 0.10692811012268066, + 0.03829650953412056, + 0.43394365906715393 + ], + "credit_map": [ + 0.030707648023962975, + 0.0064856684766709805, + 0.02579588070511818, + 0.006121575832366943, + 0.00873430073261261, + 0.014342575334012508, + 0.011039352044463158, + 0.029294362291693687, + 0.003827831242233515, + 0.0019710378255695105, + 0.004881523549556732, + 0.0031786994077265263, + 0.002681309124454856, + 0.002904767170548439, + 0.002945003332570195, + 0.004223564174026251, + 0.0034533548168838024, + 0.0008916333317756653, + 0.0029837593901902437, + 0.001260138931684196, + 0.0013668137835338712, + 0.001578310038894415, + 0.0021188915707170963, + 0.002947887172922492, + 0.0072281197644770145, + 0.0025798671413213015, + 0.0054019358940422535, + 0.002795113017782569, + 0.004396792966872454, + 0.0032766840886324644, + 0.004229470621794462, + 0.008163204416632652, + 0.019681518897414207, + 0.0027837196830660105, + 0.004622526001185179, + 0.0020483466796576977, + 0.0029067674186080694, + 0.012607394717633724, + 0.004193627741187811, + 0.03502790629863739, + 0.003774486016482115, + 0.0016851307591423392, + 0.002370483009144664, + 0.0017917697550728917, + 0.0023902026005089283, + 0.0029064284171909094, + 0.005137469153851271, + 0.004006662406027317, + 0.03670071065425873, + 0.0030012307688593864, + 0.00313559896312654, + 0.0030960829462856054, + 0.004586571827530861, + 0.04078079015016556, + 0.0037523689679801464, + 0.0040368675254285336, + 0.03546160086989403, + 0.0037586791440844536, + 0.0049202353693544865, + 0.004232513718307018, + 0.007458804175257683, + 0.03823716193437576, + 0.007837547920644283, + 0.0050215767696499825, + 0.013741171918809414, + 0.0012424223823472857, + 0.0018782074330374599, + 0.0010293612722307444, + 0.001430358155630529, + 0.014990090392529964, + 0.0018507386557757854, + 0.0021341564133763313, + 0.17512640357017517, + 0.010560897178947926, + 0.011859746649861336, + 0.00651328731328249, + 0.01109987124800682, + 0.19239936769008636, + 0.011284164153039455, + 0.015099884942173958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.029803529381752014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013371107634156942 + ], + "opposite_logprob_tokens": [ + -6.626337051391602 + ], + "logprob_reward": -0.0013371107634156942, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013371107634156942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04822147637605667, + 0.044704075902700424, + 0.010313012637197971, + 0.014178716577589512, + 0.02827143296599388, + 0.016227489337325096, + 0.055266816169023514, + 0.04415501281619072, + 0.1653931587934494, + 0.5732687711715698 + ], + "credit_map": [ + 0.009236819110810757, + 0.0025553207378834486, + 0.010822602547705173, + 0.0029736333526670933, + 0.0033471989445388317, + 0.005925190169364214, + 0.0041620731353759766, + 0.009198636747896671, + 0.007163110189139843, + 0.004661311395466328, + 0.004888797644525766, + 0.004609676077961922, + 0.007071969099342823, + 0.004760007839649916, + 0.005673472303897142, + 0.005875732284039259, + 0.0016017319867387414, + 0.0006355377845466137, + 0.0012685043038800359, + 0.0006947466754354537, + 0.0015395962400361896, + 0.0013767974451184273, + 0.001235879142768681, + 0.0019602193497121334, + 0.0018619813490658998, + 0.0014910658355802298, + 0.00202065147459507, + 0.0010846748482435942, + 0.0017061056569218636, + 0.0016758504789322615, + 0.0019991621375083923, + 0.0023392261937260628, + 0.004902913235127926, + 0.0007525277906097472, + 0.002218999434262514, + 0.00087572360644117, + 0.0008328668773174286, + 0.0058456966653466225, + 0.0014489474706351757, + 0.011393757537007332, + 0.0034531685523688793, + 0.0008122126455418766, + 0.001767619396559894, + 0.001430268632248044, + 0.0008505012956447899, + 0.0018778571393340826, + 0.0028819178696721792, + 0.0031539436895400286, + 0.01886777952313423, + 0.00224063778296113, + 0.0033401327673345804, + 0.0020743156783282757, + 0.003322037635371089, + 0.02038814127445221, + 0.002225227188318968, + 0.0028085382655262947, + 0.01612120307981968, + 0.0014588598860427737, + 0.0020835900213569403, + 0.0012958719162270427, + 0.0015421313000842929, + 0.0173743087798357, + 0.002123017329722643, + 0.002156029688194394, + 0.06526124477386475, + 0.004280933644622564, + 0.005047143902629614, + 0.002745111472904682, + 0.004818154964596033, + 0.07163333147764206, + 0.004934333730489016, + 0.0066729080863296986, + 0.23074664175510406, + 0.013943850994110107, + 0.01597951166331768, + 0.008509354665875435, + 0.014672908000648022, + 0.2542603611946106, + 0.015185913071036339, + 0.019970223307609558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.0013371107634156942 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level extreme close-up captures a classic, clear pint glass placed directly on the smooth, dark mahogany surface of a bar. The glass is filled with rich, amber-colored malt liquor, and a thick, creamy, off-white foam pours relentlessly, rising to the rim and beginning its slow, silent descent. The foam cascades down the curved sides of the glass in a thick stream of bubbles, moving at a steady, deliberate pace until it gathers in an ever-expanding, lace-like pool on the polished wood below. The entire scene is dramatically illuminated from outside the frame by a warm, low-hanging light source, casting a focused glow that accentuates the glistening, damp texture of the foam and creates bright, sharp reflections on the bar's surface, while plunging the softly blurred background into deep shadow. The atmosphere is one of quiet, lonely neglect, a wasted yet mesmerizingly beautiful moment captured in a tightly framed, intimate shot.", + "seed": 28435, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_11/index_5572_seed_28435_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0074313563294708725, + "vlm_logprob": -0.0074313563294708725, + "vlm_yes_no_margin": 5.374999952316284, + "vlm_token_credit": -0.0074313563294708725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006392862647771835 + ], + "opposite_logprob_tokens": [ + -5.756392955780029 + ], + "logprob_reward": -0.006392862647771835, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.006392862647771835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12196842581033707, + 0.31819790601730347, + 0.10484442859888077, + 0.08752201497554779, + 0.07174873352050781, + 0.06064971536397934, + 0.10409178584814072, + 0.058189988136291504, + 0.024020783603191376, + 0.048766255378723145 + ], + "credit_map": [ + 0.010310027748346329, + 0.0084682647138834, + 0.014443716034293175, + 0.013924301601946354, + 0.01343169342726469, + 0.019407253712415695, + 0.015057634562253952, + 0.026925528421998024, + 0.02790101431310177, + 0.028361164033412933, + 0.03575713559985161, + 0.027506865561008453, + 0.03366980701684952, + 0.05305190756917, + 0.033586349338293076, + 0.07836364954710007, + 0.01044350117444992, + 0.018211383372545242, + 0.006901641841977835, + 0.010697557590901852, + 0.02037455514073372, + 0.01519260834902525, + 0.01004955917596817, + 0.012973611243069172, + 0.008542032912373543, + 0.010014938190579414, + 0.0097171850502491, + 0.014925687573850155, + 0.015447281301021576, + 0.010419669561088085, + 0.011142016388475895, + 0.007313202600926161, + 0.007483429741114378, + 0.010245372541248798, + 0.007342433091253042, + 0.009132606908679008, + 0.012920031324028969, + 0.009827273897826672, + 0.0071968515403568745, + 0.007600733079016209, + 0.00557164428755641, + 0.007074721157550812, + 0.007469204720109701, + 0.011889277026057243, + 0.00717067951336503, + 0.007108388468623161, + 0.009759359061717987, + 0.00460643507540226, + 0.009003720246255398, + 0.010831822641193867, + 0.015232912264764309, + 0.013961431570351124, + 0.00994190201163292, + 0.014614305458962917, + 0.015309534966945648, + 0.015196148306131363, + 0.004001162946224213, + 0.0071817608550190926, + 0.008963535539805889, + 0.004995071329176426, + 0.008574512787163258, + 0.00898338295519352, + 0.0058564371429383755, + 0.009634124115109444, + 0.0020505161955952644, + 0.004065569955855608, + 0.0029928802978247404, + 0.002792433835566044, + 0.0026918333023786545, + 0.0034033285919576883, + 0.0031464577186852694, + 0.002877763006836176, + 0.004554028622806072, + 0.007385141681879759, + 0.004667454399168491, + 0.007429513148963451, + 0.008135098032653332, + 0.006747860927134752, + 0.005275320261716843, + 0.004571838304400444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.006392862647771835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006749332416802645 + ], + "opposite_logprob_tokens": [ + -5.006749153137207 + ], + "logprob_reward": -0.006749332416802645, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006749332416802645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09476631134748459, + 0.38614898920059204, + 0.05230306088924408, + 0.21521823108196259, + 0.0846899002790451, + 0.04688016697764397, + 0.05479460209608078, + 0.02369610406458378, + 0.02074817381799221, + 0.020754383876919746 + ], + "credit_map": [ + 0.006952625699341297, + 0.011087574996054173, + 0.009740478359162807, + 0.016503237187862396, + 0.008998317644000053, + 0.015286091715097427, + 0.012195775285363197, + 0.014002210460603237, + 0.028814291581511497, + 0.03383522853255272, + 0.04155026748776436, + 0.030855577439069748, + 0.043085455894470215, + 0.06714370101690292, + 0.039031948894262314, + 0.10183250904083252, + 0.00462842034175992, + 0.006664756685495377, + 0.006862950045615435, + 0.005017501767724752, + 0.007765911519527435, + 0.006826198659837246, + 0.007389020640403032, + 0.007148300297558308, + 0.019387440755963326, + 0.0343971848487854, + 0.010114896111190319, + 0.024614455178380013, + 0.04754248633980751, + 0.030076144263148308, + 0.029279237613081932, + 0.0198063962161541, + 0.008028460666537285, + 0.01198446936905384, + 0.0068268063478171825, + 0.012897106818854809, + 0.01342410035431385, + 0.011983677744865417, + 0.008108332753181458, + 0.0114369448274374, + 0.00395997753366828, + 0.005854558665305376, + 0.004450004547834396, + 0.007241107057780027, + 0.007057196460664272, + 0.006513175088912249, + 0.005421927664428949, + 0.006382219027727842, + 0.004968044348061085, + 0.006590838544070721, + 0.008765613660216331, + 0.00559858325868845, + 0.004199156537652016, + 0.007854162715375423, + 0.007085056975483894, + 0.00973314419388771, + 0.0019769694190472364, + 0.0038948950823396444, + 0.002423820085823536, + 0.0026849335990846157, + 0.0042284890078008175, + 0.0035528279840946198, + 0.002973199589177966, + 0.00196096976287663, + 0.0015531611861661077, + 0.0034739018883556128, + 0.002467152662575245, + 0.002196401124820113, + 0.002795169362798333, + 0.0026078769005835056, + 0.003318659495562315, + 0.0023358496837317944, + 0.002068442292511463, + 0.0031161620281636715, + 0.0024597111623734236, + 0.0023005118127912283, + 0.0030033430084586143, + 0.003001325996592641, + 0.002244015224277973, + 0.002560872118920088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.006749332416802645 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819339208304882 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.01819339208304882, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819339208304882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08495574444532394, + 0.21475660800933838, + 0.03481599688529968, + 0.22693577408790588, + 0.26254352927207947, + 0.054566796869039536, + 0.02402985841035843, + 0.041126642376184464, + 0.04323545843362808, + 0.013033632189035416 + ], + "credit_map": [ + 0.005789727903902531, + 0.008753384463489056, + 0.006283540744334459, + 0.011614846996963024, + 0.00947111938148737, + 0.012476939707994461, + 0.016640979796648026, + 0.013925212435424328, + 0.01665681041777134, + 0.017307864502072334, + 0.022915897890925407, + 0.016824733465909958, + 0.022134199738502502, + 0.03841150552034378, + 0.020846884697675705, + 0.0596587210893631, + 0.0032505353447049856, + 0.003386623691767454, + 0.0045244405046105385, + 0.006242211442440748, + 0.0032680872827768326, + 0.00467096408829093, + 0.004628320224583149, + 0.0048448145389556885, + 0.021849961951375008, + 0.03463153913617134, + 0.011241022497415543, + 0.0268630962818861, + 0.049426816403865814, + 0.0306854210793972, + 0.03281768783926964, + 0.019420228898525238, + 0.02377774938941002, + 0.04297630861401558, + 0.01575363054871559, + 0.03798982873558998, + 0.052083179354667664, + 0.03568107262253761, + 0.032584067434072495, + 0.021697677671909332, + 0.004593918099999428, + 0.008057164028286934, + 0.0046318452805280685, + 0.006300465669482946, + 0.009610037319362164, + 0.007645494304597378, + 0.006658453494310379, + 0.007069414481520653, + 0.0026969341561198235, + 0.0029142219573259354, + 0.003722645342350006, + 0.002130283508449793, + 0.0023489443119615316, + 0.003543224884197116, + 0.0028868941590189934, + 0.003786709625273943, + 0.003167614573612809, + 0.006405131425708532, + 0.0044855764135718346, + 0.005302838049829006, + 0.008419076912105083, + 0.00564458267763257, + 0.0047143627889454365, + 0.0029874606989324093, + 0.002968928776681423, + 0.007655453402549028, + 0.0053034559823572636, + 0.0054221441969275475, + 0.0056463442742824554, + 0.005983331706374884, + 0.006374236661940813, + 0.003881567157804966, + 0.0016245184233412147, + 0.0021040327847003937, + 0.001215465716086328, + 0.001414091675542295, + 0.0015764483250677586, + 0.001989765092730522, + 0.0015401631826534867, + 0.0015691471053287387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.01819339208304882 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639456048607826 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639456048607826, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639456048607826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14183185994625092, + 0.4960346221923828, + 0.060983870178461075, + 0.06733831018209457, + 0.045475881546735764, + 0.08878646790981293, + 0.06334350258111954, + 0.017309486865997314, + 0.008750383742153645, + 0.010145586915314198 + ], + "credit_map": [ + 0.011827550828456879, + 0.01260026078671217, + 0.01580963470041752, + 0.013658178970217705, + 0.019825220108032227, + 0.02169038914144039, + 0.01706700585782528, + 0.029353616759181023, + 0.04093306139111519, + 0.04348974674940109, + 0.05406138673424721, + 0.039959777146577835, + 0.05252298340201378, + 0.0843401551246643, + 0.05326602980494499, + 0.1274615079164505, + 0.004461326636373997, + 0.008508327417075634, + 0.0036725630052387714, + 0.004679512232542038, + 0.011665699072182178, + 0.009123648516833782, + 0.009811659343540668, + 0.009061132557690144, + 0.0066339378245174885, + 0.009291871450841427, + 0.005744937341660261, + 0.00746770016849041, + 0.014240292832255363, + 0.00892503373324871, + 0.00936032459139824, + 0.005674211774021387, + 0.004982935730367899, + 0.007931159809231758, + 0.003321875585243106, + 0.0053632534109056, + 0.008769243024289608, + 0.006386473309248686, + 0.0051291570998728275, + 0.003591781249269843, + 0.008787527680397034, + 0.00940769538283348, + 0.013181768357753754, + 0.01618798077106476, + 0.008421670645475388, + 0.011672554537653923, + 0.008736123330891132, + 0.012391150929033756, + 0.006828624755144119, + 0.0073457988910377026, + 0.01160014420747757, + 0.005356849171221256, + 0.005715252831578255, + 0.009644006378948689, + 0.006140960846096277, + 0.010711867362260818, + 0.001686671981588006, + 0.0024509751237928867, + 0.00221438379958272, + 0.0019389089429751039, + 0.0024741291999816895, + 0.0023477349895983934, + 0.0018473172094672918, + 0.0023493655025959015, + 0.0009662167867645621, + 0.0017520596738904715, + 0.000797624175902456, + 0.0009896029951050878, + 0.0013081926153972745, + 0.001305649639107287, + 0.0009105553035624325, + 0.0007204823778010905, + 0.0011416299967095256, + 0.0012834775261580944, + 0.0013394041452556849, + 0.0008991825161501765, + 0.0015368475578725338, + 0.00141617760527879, + 0.0011901712277904153, + 0.0013386965729296207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.004639456048607826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011817384511232376 + ], + "opposite_logprob_tokens": [ + -6.751181602478027 + ], + "logprob_reward": -0.0011817384511232376, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011817384511232376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0697416365146637, + 0.47004014253616333, + 0.20151063799858093, + 0.0311397984623909, + 0.0723954290151596, + 0.04372341185808182, + 0.04324076324701309, + 0.030313408002257347, + 0.021342378109693527, + 0.016552459448575974 + ], + "credit_map": [ + 0.005443007685244083, + 0.006392570212483406, + 0.0072967857122421265, + 0.006818958558142185, + 0.007889563217759132, + 0.011932591907680035, + 0.00748114287853241, + 0.016487011685967445, + 0.0373854786157608, + 0.040744341909885406, + 0.05048222094774246, + 0.037340521812438965, + 0.05123252421617508, + 0.08111122995615005, + 0.04863671585917473, + 0.12310709059238434, + 0.01785002462565899, + 0.03481808304786682, + 0.016752658411860466, + 0.01825374737381935, + 0.044242240488529205, + 0.028759686276316643, + 0.019847512245178223, + 0.020986665040254593, + 0.0029253214597702026, + 0.002825153525918722, + 0.004038628656417131, + 0.006647326983511448, + 0.003348547499626875, + 0.003879090305417776, + 0.00351540045812726, + 0.003960326313972473, + 0.006842086091637611, + 0.01187572255730629, + 0.004827519878745079, + 0.010402048006653786, + 0.013793268240988255, + 0.010004675015807152, + 0.00919261109083891, + 0.005457495339214802, + 0.005262571386992931, + 0.006246485281735659, + 0.004840319510549307, + 0.005482711363583803, + 0.006265558302402496, + 0.005986636504530907, + 0.004806599114090204, + 0.0048325262032449245, + 0.004694317001849413, + 0.00457599526271224, + 0.007548287510871887, + 0.005190849769860506, + 0.005057761445641518, + 0.005866208113729954, + 0.005648081190884113, + 0.004659258760511875, + 0.0032841258216649294, + 0.005038926377892494, + 0.002657542238011956, + 0.003918835893273354, + 0.005125236697494984, + 0.0038846503011882305, + 0.0034676766954362392, + 0.0029364097863435745, + 0.002434342633932829, + 0.003529037581756711, + 0.003053468419238925, + 0.001955263549461961, + 0.0024958758149296045, + 0.0033005173318088055, + 0.002088212175294757, + 0.0024856592062860727, + 0.0020124956499785185, + 0.002322954358533025, + 0.00213590613566339, + 0.0016201926628127694, + 0.002285548485815525, + 0.002384894760325551, + 0.0020989326294511557, + 0.001691534067504108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.0011817384511232376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level extreme close-up captures a classic, clear pint glass placed directly on the smooth, dark mahogany surface of a bar. The glass is filled with rich, amber-colored malt liquor, and a thick, creamy, off-white foam pours relentlessly, rising to the rim and beginning its slow, silent descent. The foam cascades down the curved sides of the glass in a thick stream of bubbles, moving at a steady, deliberate pace until it gathers in an ever-expanding, lace-like pool on the polished wood below. The entire scene is dramatically illuminated from outside the frame by a warm, low-hanging light source, casting a focused glow that accentuates the glistening, damp texture of the foam and creates bright, sharp reflections on the bar's surface, while plunging the softly blurred background into deep shadow. The atmosphere is one of quiet, lonely neglect, a wasted yet mesmerizingly beautiful moment captured in a tightly framed, intimate shot.", + "seed": 28436, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_11/index_5572_seed_28436_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025018238741904498, + "vlm_logprob": -0.025018238741904498, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.025018238741904498, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013589724898338318 + ], + "opposite_logprob_tokens": [ + -5.138589859008789 + ], + "logprob_reward": -0.013589724898338318, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.013589724898338318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046344880014657974, + 0.8074414730072021, + 0.011368121951818466, + 0.048791706562042236, + 0.058219049125909805, + 0.009232846088707447, + 0.005750807002186775, + 0.002755609108135104, + 0.004449945408850908, + 0.005645647644996643 + ], + "credit_map": [ + 0.0029506543651223183, + 0.001250519067980349, + 0.003258913056924939, + 0.002710289554670453, + 0.0021983024198561907, + 0.01516704075038433, + 0.004015824291855097, + 0.014793330803513527, + 0.03395228087902069, + 0.005187763832509518, + 0.018596110865473747, + 0.019976848736405373, + 0.010707749985158443, + 0.3475348651409149, + 0.018374662846326828, + 0.35311105847358704, + 0.0017507831798866391, + 0.000593912205658853, + 0.0019621429964900017, + 0.00179688585922122, + 0.0006712532485835254, + 0.0017537466483190656, + 0.001099894056096673, + 0.0017395021859556437, + 0.005087157245725393, + 0.0012587819946929812, + 0.0038833029102534056, + 0.003971016500145197, + 0.0024319635704159737, + 0.011740287765860558, + 0.0083667803555727, + 0.01205240935087204, + 0.0025304295122623444, + 0.0004092634189873934, + 0.0013287065085023642, + 0.0009732540929690003, + 0.0006289008306339383, + 0.025629203766584396, + 0.0012088860385119915, + 0.025510400533676147, + 0.001523595885373652, + 0.00048402207903563976, + 0.0009035986149683595, + 0.0014724410139024258, + 0.0006152227870188653, + 0.0010835183784365654, + 0.001782127539627254, + 0.0013683177530765533, + 0.0008105749730020761, + 0.000400054850615561, + 0.0010544491233304143, + 0.0007151958998292685, + 0.0005760716157965362, + 0.0007136821513995528, + 0.0009051859378814697, + 0.0005755922175012529, + 0.000400654214899987, + 0.00018788262968882918, + 0.00033658475149422884, + 0.0002918875543400645, + 0.00030051375506445765, + 0.0003859321877826005, + 0.0005305778468027711, + 0.00032157590612769127, + 0.0005035314243286848, + 0.00029020890360698104, + 0.0003310278116259724, + 0.0004400268953759223, + 0.0006230081198737025, + 0.0005316759925335646, + 0.0012906371848657727, + 0.0004398280871100724, + 0.0006627158145420253, + 0.0006403494044207036, + 0.000555652950424701, + 0.0005334377055987716, + 0.0011219251900911331, + 0.0006523024057969451, + 0.0011200925800949335, + 0.0003591712156776339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.013589724898338318 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214809983968735 + ], + "opposite_logprob_tokens": [ + -4.264214992523193 + ], + "logprob_reward": -0.014214809983968735, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214809983968735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051879435777664185, + 0.48663219809532166, + 0.05802997574210167, + 0.05338119715452194, + 0.29466715455055237, + 0.010202025063335896, + 0.018461553379893303, + 0.00883591640740633, + 0.008708160370588303, + 0.009202348068356514 + ], + "credit_map": [ + 0.005906181409955025, + 0.0032046926207840443, + 0.0034896088764071465, + 0.004788518417626619, + 0.00543493777513504, + 0.010850805789232254, + 0.007476795930415392, + 0.010727898217737675, + 0.020839910954236984, + 0.003421356901526451, + 0.011427282355725765, + 0.012159730307757854, + 0.0073621696792542934, + 0.2090095728635788, + 0.011473395861685276, + 0.21093876659870148, + 0.00365820131264627, + 0.00190923479385674, + 0.0029704708140343428, + 0.003092375583946705, + 0.0029774170834571123, + 0.019297735765576363, + 0.005095100030303001, + 0.019029438495635986, + 0.0035767590161412954, + 0.0013911753194406629, + 0.003168438794091344, + 0.005122134927660227, + 0.003521570237353444, + 0.010913800448179245, + 0.013655662536621094, + 0.012031657621264458, + 0.011864149942994118, + 0.0018789139576256275, + 0.006107853259891272, + 0.004972422029823065, + 0.0030249899718910456, + 0.13070444762706757, + 0.0054827965795993805, + 0.13063156604766846, + 0.0013634110800921917, + 0.0006617350736632943, + 0.0009460636647418141, + 0.0009483401081524789, + 0.001095217652618885, + 0.0013821538304910064, + 0.002064290689304471, + 0.001740812323987484, + 0.0030539047438651323, + 0.0010049381526187062, + 0.00220997235737741, + 0.001447197631932795, + 0.0014120135456323624, + 0.002806443953886628, + 0.0022043175995349884, + 0.004322764929383993, + 0.0014847335405647755, + 0.0006648861453868449, + 0.0009716960485093296, + 0.0006687699351459742, + 0.0012409697519615293, + 0.0011986736208200455, + 0.001619565300643444, + 0.0009866226464509964, + 0.0008971852948889136, + 0.0007313791429623961, + 0.0011454252526164055, + 0.001122887129895389, + 0.0012273471802473068, + 0.0010358517756685615, + 0.0015606944216415286, + 0.0009873901726678014, + 0.001004036865197122, + 0.000724580604583025, + 0.0012893902603536844, + 0.0009404551819898188, + 0.0014005698030814528, + 0.0011020461097359657, + 0.002102338708937168, + 0.0006389300688169897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.014214809983968735 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901103049516678 + ], + "opposite_logprob_tokens": [ + -2.5790109634399414 + ], + "logprob_reward": -0.07901103049516678, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901103049516678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054296523332595825, + 0.617796003818512, + 0.1413721889257431, + 0.051111310720443726, + 0.08560820668935776, + 0.008262558840215206, + 0.013402115553617477, + 0.012733684852719307, + 0.007912063971161842, + 0.007505329325795174 + ], + "credit_map": [ + 0.00505299586802721, + 0.001829601125791669, + 0.0030168937519192696, + 0.0033360717352479696, + 0.0026712820399552584, + 0.01710319332778454, + 0.004462788812816143, + 0.016823695972561836, + 0.02629593200981617, + 0.00416537094861269, + 0.01415594108402729, + 0.015133817680180073, + 0.008730483241379261, + 0.2658293545246124, + 0.014471003785729408, + 0.2690140902996063, + 0.008039429783821106, + 0.0016824326012283564, + 0.003910123836249113, + 0.0041152965277433395, + 0.0022598293144255877, + 0.05896842107176781, + 0.0049773091450333595, + 0.05741935223340988, + 0.003172188764438033, + 0.0013054549926891923, + 0.002836533822119236, + 0.004869486205279827, + 0.0035223844461143017, + 0.010221394710242748, + 0.013754465617239475, + 0.011429401114583015, + 0.0032127697486430407, + 0.0006953621050342917, + 0.001957717351615429, + 0.0017427204875275493, + 0.0013505240203812718, + 0.037222929298877716, + 0.0022039180621504784, + 0.03722226247191429, + 0.0009217890328727663, + 0.0005191212985664606, + 0.0008307679672725499, + 0.0007678618421778083, + 0.001262263860553503, + 0.0011328079272061586, + 0.0013638209784403443, + 0.0014641251182183623, + 0.0017158366972580552, + 0.0009403863223269582, + 0.001836911658756435, + 0.0012270711595192552, + 0.001544735860079527, + 0.0018154468853026628, + 0.0019740944262593985, + 0.0023476325441151857, + 0.0015353646595031023, + 0.0007655113586224616, + 0.0014513969654217362, + 0.0013322102604433894, + 0.0015547063667327166, + 0.001647388911806047, + 0.0030522269662469625, + 0.001394879654981196, + 0.0008109683403745294, + 0.0005665652570314705, + 0.0006556800799444318, + 0.0008747957181185484, + 0.001209916197694838, + 0.0008734774892218411, + 0.0021023445297032595, + 0.0008183160680346191, + 0.0008452378679066896, + 0.0005982774309813976, + 0.0008713632705621421, + 0.0008330443524755538, + 0.001242625992745161, + 0.0008582026930525899, + 0.001604054938070476, + 0.0006525226635858417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.07901103049516678 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016074350103735924 + ], + "opposite_logprob_tokens": [ + -4.141074180603027 + ], + "logprob_reward": -0.016074350103735924, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016074350103735924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021527428179979324, + 0.4240821897983551, + 0.07812938839197159, + 0.013249770738184452, + 0.4484545588493347, + 0.00236709788441658, + 0.005995075684040785, + 0.0029380323830991983, + 0.0017683458281680942, + 0.0014881716342642903 + ], + "credit_map": [ + 0.0050103445537388325, + 0.0009580656769685447, + 0.003030999330803752, + 0.001951392856426537, + 0.0013188906013965607, + 0.0038904619868844748, + 0.002226645592600107, + 0.003140626708045602, + 0.017973387613892555, + 0.0028446975629776716, + 0.00978105142712593, + 0.010381062515079975, + 0.005616946145892143, + 0.18243855237960815, + 0.009785974398255348, + 0.18526047468185425, + 0.004576972685754299, + 0.0008672430994920433, + 0.0033838616218417883, + 0.002898093778640032, + 0.001269355183467269, + 0.03167742118239403, + 0.002610768424347043, + 0.03084567002952099, + 0.0014083090936765075, + 0.00036389988963492215, + 0.001017349073663354, + 0.000996666494756937, + 0.0007922480581328273, + 0.0027691347058862448, + 0.002894517732784152, + 0.003007645485922694, + 0.018532095476984978, + 0.0029241801239550114, + 0.009393219836056232, + 0.007402127608656883, + 0.004612616263329983, + 0.19853276014328003, + 0.008164173923432827, + 0.19889333844184875, + 0.0003566702944226563, + 0.0001376058644382283, + 0.00041641300776973367, + 0.0002965768799185753, + 0.00019284898007754236, + 0.0002698588650673628, + 0.0003492625255603343, + 0.0003478613798506558, + 0.00098852114751935, + 0.00031026737997308373, + 0.0008172820089384913, + 0.0005658758454956114, + 0.00038248547934927046, + 0.0008691335679031909, + 0.0008271588012576103, + 0.0012343511916697025, + 0.00038550954195670784, + 0.00018502442981116474, + 0.0002523143484722823, + 0.00030950328800827265, + 0.00030296522891148925, + 0.00037584875826723874, + 0.0007826286018826067, + 0.00034423801116645336, + 0.00021222676150500774, + 0.00012931309174746275, + 0.00019745920144487172, + 0.0002008177834795788, + 0.0002702903875615448, + 0.00018546961655374616, + 0.000413276458857581, + 0.00015949239605106413, + 0.00019905538647435606, + 0.00010113006283063442, + 0.00011305739462841302, + 0.0001681641733739525, + 0.0002479207469150424, + 0.00019358936697244644, + 0.00034570670686662197, + 0.00011954778165090829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.016074350103735924 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a classic pint glass placed on a dark mahogany bar surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the glass filled with amber-colored malt liquor and a thick, off-white foam pouring over the rim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the foam cascade down the sides of the glass in a thick stream of bubbles at a steady pace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm, low-hanging light source from outside the frame creating sharp reflections on the bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred and plunged into deep shadow while the foreground is in focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201278228312731 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.002201278228312731, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201278228312731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18187199532985687, + 0.41096410155296326, + 0.12793244421482086, + 0.0298434030264616, + 0.1650296151638031, + 0.023138193413615227, + 0.015338568016886711, + 0.02278013341128826, + 0.01065407320857048, + 0.012447542510926723 + ], + "credit_map": [ + 0.01474787574261427, + 0.004365939646959305, + 0.014944946393370628, + 0.008857913315296173, + 0.006517773028463125, + 0.059420906007289886, + 0.011568918824195862, + 0.061447709798812866, + 0.017286675050854683, + 0.0032037256751209497, + 0.011466996744275093, + 0.010412466712296009, + 0.005717368330806494, + 0.17462588846683502, + 0.0093497633934021, + 0.17890116572380066, + 0.010287368670105934, + 0.002607781905680895, + 0.009822260588407516, + 0.0074689933098852634, + 0.0029193577356636524, + 0.04595346003770828, + 0.00505071971565485, + 0.04382249340415001, + 0.003956736531108618, + 0.0011983317090198398, + 0.002816353924572468, + 0.0032512519974261522, + 0.002229077974334359, + 0.004485608544200659, + 0.0071632456965744495, + 0.004742794670164585, + 0.006835055071860552, + 0.0013951974688097835, + 0.004021838307380676, + 0.0033258553594350815, + 0.002514774212613702, + 0.07191503793001175, + 0.003446131246164441, + 0.07157568633556366, + 0.004288237076252699, + 0.0011226576752960682, + 0.0025338924024254084, + 0.002756659174337983, + 0.002761504612863064, + 0.002643126994371414, + 0.003204327542334795, + 0.003827787470072508, + 0.0016478204634040594, + 0.001200408674776554, + 0.0018214076990261674, + 0.001924901851452887, + 0.0019987146370112896, + 0.001859453390352428, + 0.0025185521226376295, + 0.0023673067335039377, + 0.002652019029483199, + 0.00132306560408324, + 0.0028291812632232904, + 0.002698622876778245, + 0.0023330356925725937, + 0.0027921146247535944, + 0.005474897567182779, + 0.002677195006981492, + 0.0011517421808093786, + 0.0008481528493575752, + 0.0010526522528380156, + 0.0012459143763408065, + 0.0016877676825970411, + 0.0012600722257047892, + 0.002422376535832882, + 0.0009853935334831476, + 0.0020646043121814728, + 0.0010183805134147406, + 0.0011838253121823072, + 0.0011362445075064898, + 0.0016401534667238593, + 0.0017783802468329668, + 0.0025120608042925596, + 0.0011138916015625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.002201278228312731 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_12/index_tensor([7091, 7091])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_12/index_tensor([7091, 7091])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a19f649aac172cbcc3bcf232119da5b8eb074a6b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_12/index_tensor([7091, 7091])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, ultra-high-definition, vertical medium shot captures a dynamic ink painting at eye level. Against a backdrop of Chinese landscapes that seem to come alive and flow gently across rice paper, a woman with a delicate face, long black hair loosely tied by an ancient jade hairpin, gracefully leans forward. She wears a flowing, pale-blue silk Hanfu dress, its wide sleeves gently falling with her movements. The lapels are embroidered with delicate, pale-ink bamboo leaf patterns. Her movements are focused and deliberate. A slender hand reaches down, carefully grasping the handle of a modern canvas diaper bag from the ancient bluestone slabs, weathered by time and dotted with moss, and then slowly lifts it. The off-white canvas bag stands in stark yet harmonious contrast to its classical surroundings. A vibrant, traditional auspicious cloud pattern is clearly embroidered on the front in delicate light-gold thread. The light is as soft and even as if filtering through thin rice paper. Without a clear source, it captures every detail, from the sheen of the Hanfu silk to the texture of the canvas to the woman's serene expression. As she rises, the distant mountains, rendered with thick and thin brushstrokes, continue to swirl with mist. The entire scene, nestled within the black, white, and gray of ink wash painting, is infused with vitality by a touch of pale blue on the sleeve and a touch of light gold on the embroidery. A unique atmosphere, a timeless, tranquil poetry, a blend of classical aesthetics and modern pragmatism, permeates the space.", + "seed": 13337, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_12/index_7091_seed_13337_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06919542541727423, + "vlm_logprob": -0.06919542541727423, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.06919542541727423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01137709990143776 + ], + "opposite_logprob_tokens": [ + -5.386377334594727 + ], + "logprob_reward": -0.01137709990143776, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01137709990143776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021427838131785393, + 0.015358035452663898, + 0.011182546615600586, + 0.017716743052005768, + 0.011672521010041237, + 0.0191781185567379, + 0.7871793508529663, + 0.043644338846206665, + 0.04373789578676224, + 0.02890254743397236 + ], + "credit_map": [ + 0.001602492993697524, + 0.003646321129053831, + 0.0036413357593119144, + 0.002456517657265067, + 0.002094530500471592, + 0.0029946910217404366, + 0.0025553409941494465, + 0.0024366071447730064, + 0.0012408810434862971, + 0.0019499522168189287, + 0.0029147148597985506, + 0.0023077703081071377, + 0.0017998941475525498, + 0.0019997297786176205, + 0.0020501550752669573, + 0.0010949387215077877, + 0.0008931987104006112, + 0.0010786459315568209, + 0.00195831130258739, + 0.0017846738919615746, + 0.0014379664789885283, + 0.0011726953089237213, + 0.0017261820612475276, + 0.0011308727553114295, + 0.00138198456261307, + 0.0025859565939754248, + 0.0019523174269124866, + 0.002249092096462846, + 0.0022183130495250225, + 0.0034188281279057264, + 0.00245825108140707, + 0.0014519987162202597, + 0.0007695947424508631, + 0.0016371365636587143, + 0.0010313504608348012, + 0.0014725541695952415, + 0.0012333242921158671, + 0.001954578561708331, + 0.002707180567085743, + 0.0008668018272146583, + 0.00196612486615777, + 0.0023407607804983854, + 0.002596429781988263, + 0.0033245685044676065, + 0.0033348931465297937, + 0.0019369354704394937, + 0.002561924746260047, + 0.0011164812603965402, + 0.028743820264935493, + 0.030585812404751778, + 0.029074987396597862, + 0.06579016149044037, + 0.21628133952617645, + 0.05915677174925804, + 0.3315214216709137, + 0.02602507919073105, + 0.004329430405050516, + 0.0048442804254591465, + 0.0032536941580474377, + 0.004427689127624035, + 0.005433693062514067, + 0.0064552221447229385, + 0.012001418508589268, + 0.0028989112470299006, + 0.0062739793211221695, + 0.006810471881181002, + 0.005411061458289623, + 0.004892970435321331, + 0.008044255897402763, + 0.004525593016296625, + 0.0036840930115431547, + 0.004095470067113638, + 0.0024004019796848297, + 0.004984704311937094, + 0.0025963520165532827, + 0.003904111450538039, + 0.003445462556555867, + 0.003824661485850811, + 0.0037036598660051823, + 0.004043195396661758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.01137709990143776 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331020124256611 + ], + "opposite_logprob_tokens": [ + -3.7733101844787598 + ], + "logprob_reward": -0.02331020124256611, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331020124256611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08076120913028717, + 0.07830541580915451, + 0.07818666100502014, + 0.03378099575638771, + 0.03098994493484497, + 0.17821821570396423, + 0.21692009270191193, + 0.060210149735212326, + 0.15504217147827148, + 0.08758516609668732 + ], + "credit_map": [ + 0.0058416821993887424, + 0.012710653245449066, + 0.01143545564264059, + 0.010275629349052906, + 0.01079932413995266, + 0.012327946722507477, + 0.008934419602155685, + 0.008436089381575584, + 0.004929039627313614, + 0.014296506531536579, + 0.007346798665821552, + 0.02123347297310829, + 0.01083475910127163, + 0.005187279544770718, + 0.011020952835679054, + 0.003456604667007923, + 0.007648777682334185, + 0.006882151588797569, + 0.01741505041718483, + 0.00873629841953516, + 0.008950000628829002, + 0.007420509587973356, + 0.010840860195457935, + 0.01029300969094038, + 0.0030310337897390127, + 0.0033061890862882137, + 0.005002949386835098, + 0.004120687488466501, + 0.003795553930103779, + 0.005004920996725559, + 0.004770851694047451, + 0.004748804960399866, + 0.0017665621126070619, + 0.0032178377732634544, + 0.004248473327606916, + 0.0032149285543709993, + 0.0033504217863082886, + 0.0058606332167983055, + 0.007011059671640396, + 0.002320027444511652, + 0.014204890467226505, + 0.03709106519818306, + 0.02116539515554905, + 0.03844015300273895, + 0.027515051886439323, + 0.0112575339153409, + 0.019899489358067513, + 0.008644632063806057, + 0.010150667279958725, + 0.009924964047968388, + 0.010715540498495102, + 0.021757157519459724, + 0.056488096714019775, + 0.013961683958768845, + 0.08432578295469284, + 0.009596181102097034, + 0.005011273082345724, + 0.008805842138826847, + 0.006489492021501064, + 0.009401889517903328, + 0.008341231383383274, + 0.01038258709013462, + 0.006749426014721394, + 0.005028406623750925, + 0.018392065539956093, + 0.018878761678934097, + 0.019956624135375023, + 0.012170487083494663, + 0.03461560234427452, + 0.017831455916166306, + 0.01536189392209053, + 0.017835266888141632, + 0.0059418813325464725, + 0.01914273388683796, + 0.00873671192675829, + 0.008951086550951004, + 0.009956474415957928, + 0.008558851666748524, + 0.010626379400491714, + 0.015671033412218094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.02331020124256611 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22556714713573456 + ], + "opposite_logprob_tokens": [ + -1.600567102432251 + ], + "logprob_reward": -0.22556714713573456, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22556714713573456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048170775175094604, + 0.013283868320286274, + 0.043389517813920975, + 0.03770618140697479, + 0.0074857971630990505, + 0.02303162030875683, + 0.5925825238227844, + 0.09549643844366074, + 0.0882522240281105, + 0.05060099437832832 + ], + "credit_map": [ + 0.005026536528021097, + 0.006668104324489832, + 0.005728261079639196, + 0.0056931921280920506, + 0.005977639928460121, + 0.008950671181082726, + 0.0056987423449754715, + 0.004427638370543718, + 0.0009450765792280436, + 0.001475437544286251, + 0.0018376177176833153, + 0.001966360490769148, + 0.0014094712678343058, + 0.0018130383687093854, + 0.0025333540979772806, + 0.0013035136507824063, + 0.0037136452738195658, + 0.005885590799152851, + 0.0076660714112222195, + 0.0052983458153903484, + 0.004153855610638857, + 0.0074051618576049805, + 0.004809061996638775, + 0.00445779412984848, + 0.004243412986397743, + 0.00331830233335495, + 0.0051913331262767315, + 0.0035721699241548777, + 0.003972044680267572, + 0.006592693738639355, + 0.004693468101322651, + 0.006122763268649578, + 0.0005005930433981121, + 0.0006981472251936793, + 0.0011881932150572538, + 0.0008567238110117614, + 0.0008638304425403476, + 0.0014373406302183867, + 0.0011228376533836126, + 0.000818132481072098, + 0.0019406946375966072, + 0.0039453390054404736, + 0.0027077198028564453, + 0.0032233484089374542, + 0.0030867252498865128, + 0.0031880391761660576, + 0.0031123869121074677, + 0.0018273730529472232, + 0.02161455899477005, + 0.023953983560204506, + 0.02102559246122837, + 0.04667649418115616, + 0.16465766727924347, + 0.04577585682272911, + 0.24952302873134613, + 0.019355403259396553, + 0.00971495546400547, + 0.009567721746861935, + 0.006028044503182173, + 0.01133011281490326, + 0.011546975001692772, + 0.013051646761596203, + 0.029253089800477028, + 0.00500390725210309, + 0.007703906390815973, + 0.010949456132948399, + 0.006567027885466814, + 0.005758398212492466, + 0.024170948192477226, + 0.012519831769168377, + 0.010489738546311855, + 0.010092932730913162, + 0.0037920631002634764, + 0.006822796072810888, + 0.005708659067749977, + 0.006874095182865858, + 0.005826376844197512, + 0.008289172314107418, + 0.007983884774148464, + 0.0053039598278701305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.22556714713573456 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898779213428497 + ], + "opposite_logprob_tokens": [ + -2.5789878368377686 + ], + "logprob_reward": -0.07898779213428497, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898779213428497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07244621217250824, + 0.028829116374254227, + 0.06399739533662796, + 0.03402227908372879, + 0.026916855946183205, + 0.04802112653851509, + 0.4085572361946106, + 0.06012594699859619, + 0.14795643091201782, + 0.10912739485502243 + ], + "credit_map": [ + 0.005115114618092775, + 0.009556052275002003, + 0.013756193220615387, + 0.009106473065912724, + 0.00797668844461441, + 0.010983566753566265, + 0.010601406916975975, + 0.00535071175545454, + 0.0023723472841084003, + 0.003264166647568345, + 0.003230196190997958, + 0.006160063203424215, + 0.003672270802780986, + 0.004749675747007132, + 0.0030734592583030462, + 0.0023069344460964203, + 0.006467356812208891, + 0.007215593475848436, + 0.010806267149746418, + 0.005144548136740923, + 0.00652466481551528, + 0.01050816010683775, + 0.006762669421732426, + 0.010568130761384964, + 0.0030385481659322977, + 0.003937011118978262, + 0.004875117912888527, + 0.003911878913640976, + 0.0034016522113233805, + 0.004339350387454033, + 0.00468445336446166, + 0.00583426421508193, + 0.0015611745184287429, + 0.001804753439500928, + 0.0029712060932070017, + 0.001992059638723731, + 0.004015491344034672, + 0.00394890271127224, + 0.008682460524141788, + 0.001940806396305561, + 0.004174661822617054, + 0.008330859243869781, + 0.006110153626650572, + 0.00799574051052332, + 0.007380479481071234, + 0.004902679473161697, + 0.005237357225269079, + 0.0038891960866749287, + 0.017043549567461014, + 0.014144022017717361, + 0.016882428899407387, + 0.042063433676958084, + 0.09201375395059586, + 0.039992544800043106, + 0.17216436564922333, + 0.014253121800720692, + 0.005532390903681517, + 0.006728304550051689, + 0.006189085077494383, + 0.007294349372386932, + 0.007038347888737917, + 0.008806838653981686, + 0.014580362476408482, + 0.003956262953579426, + 0.012668369337916374, + 0.020304696634411812, + 0.008182425051927567, + 0.00920067261904478, + 0.03589952737092972, + 0.020299190655350685, + 0.022003261372447014, + 0.0193982794880867, + 0.009491239674389362, + 0.024096865206956863, + 0.012061584740877151, + 0.011797755025327206, + 0.009979977272450924, + 0.010533960536122322, + 0.01171494647860527, + 0.01945105381309986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.07898779213428497 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734886672347784 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734886672347784, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734886672347784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04120452702045441, + 0.010883788578212261, + 0.009439903311431408, + 0.02023394964635372, + 0.008710269816219807, + 0.014722771942615509, + 0.7865904569625854, + 0.030954301357269287, + 0.053159199655056, + 0.02410084754228592 + ], + "credit_map": [ + 0.0034363612066954374, + 0.005355640780180693, + 0.005186338443309069, + 0.005363296251744032, + 0.005618780851364136, + 0.006380727048963308, + 0.004813318606466055, + 0.005050065461546183, + 0.0009012444061227143, + 0.0012390550691634417, + 0.0021529332734644413, + 0.0012489974033087492, + 0.0010938021587207913, + 0.0018746505957096815, + 0.0012790312757715583, + 0.00109407480340451, + 0.0009811694035306573, + 0.001196148688904941, + 0.0016238406533375382, + 0.0009687449783086777, + 0.0010096895275637507, + 0.0014957193052396178, + 0.0011029924498870969, + 0.0010615979554131627, + 0.001842385157942772, + 0.002673767041414976, + 0.002390272682532668, + 0.0021945969201624393, + 0.0023254575207829475, + 0.0037369108758866787, + 0.002721899887546897, + 0.0023486600257456303, + 0.0006923599867150187, + 0.0007694917148910463, + 0.0012385068694129586, + 0.0008381892112083733, + 0.0010322954040020704, + 0.0017105020815506577, + 0.0014123378787189722, + 0.001016586902551353, + 0.0015357326483353972, + 0.0013437146553769708, + 0.002163916127756238, + 0.002780310343950987, + 0.002329193986952305, + 0.001605593366548419, + 0.0015444520395249128, + 0.0014198586577549577, + 0.029126573354005814, + 0.031085729598999023, + 0.02925121784210205, + 0.06381409615278244, + 0.21835225820541382, + 0.05928615480661392, + 0.3295418918132782, + 0.02613254450261593, + 0.0027587306685745716, + 0.002940779086202383, + 0.002695902716368437, + 0.0043163481168448925, + 0.0049584065563976765, + 0.004713178612291813, + 0.006653499323874712, + 0.0019174537155777216, + 0.008664080873131752, + 0.007698811590671539, + 0.007145405281335115, + 0.006111785303801298, + 0.007554186508059502, + 0.005198095925152302, + 0.004222759511321783, + 0.00656407605856657, + 0.002298736246302724, + 0.00423955824226141, + 0.0031605465337634087, + 0.004178657662123442, + 0.0020375228486955166, + 0.002641060622408986, + 0.002526597585529089, + 0.0030181673355400562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.006734886672347784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, ultra-high-definition, vertical medium shot captures a dynamic ink painting at eye level. Against a backdrop of Chinese landscapes that seem to come alive and flow gently across rice paper, a woman with a delicate face, long black hair loosely tied by an ancient jade hairpin, gracefully leans forward. She wears a flowing, pale-blue silk Hanfu dress, its wide sleeves gently falling with her movements. The lapels are embroidered with delicate, pale-ink bamboo leaf patterns. Her movements are focused and deliberate. A slender hand reaches down, carefully grasping the handle of a modern canvas diaper bag from the ancient bluestone slabs, weathered by time and dotted with moss, and then slowly lifts it. The off-white canvas bag stands in stark yet harmonious contrast to its classical surroundings. A vibrant, traditional auspicious cloud pattern is clearly embroidered on the front in delicate light-gold thread. The light is as soft and even as if filtering through thin rice paper. Without a clear source, it captures every detail, from the sheen of the Hanfu silk to the texture of the canvas to the woman's serene expression. As she rises, the distant mountains, rendered with thick and thin brushstrokes, continue to swirl with mist. The entire scene, nestled within the black, white, and gray of ink wash painting, is infused with vitality by a touch of pale blue on the sleeve and a touch of light gold on the embroidery. A unique atmosphere, a timeless, tranquil poetry, a blend of classical aesthetics and modern pragmatism, permeates the space.", + "seed": 13338, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_12/index_7091_seed_13338_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03671941608190536, + "vlm_logprob": -0.03671941608190536, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03671941608190536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0153699005022645 + ], + "opposite_logprob_tokens": [ + -4.765369892120361 + ], + "logprob_reward": -0.0153699005022645, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0153699005022645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08752509206533432, + 0.044941145926713943, + 0.057276494801044464, + 0.06380219757556915, + 0.06698533892631531, + 0.14358168840408325, + 0.35598522424697876, + 0.07737711071968079, + 0.04717414453625679, + 0.055351607501506805 + ], + "credit_map": [ + 0.01146619115024805, + 0.011868707835674286, + 0.00832359492778778, + 0.011995978653430939, + 0.01580948382616043, + 0.008617658168077469, + 0.009761828929185867, + 0.009681645780801773, + 0.005495536606758833, + 0.003950778394937515, + 0.003389607882127166, + 0.004685889929533005, + 0.00727536017075181, + 0.006987988483160734, + 0.005605294369161129, + 0.007550687063485384, + 0.004765807650983334, + 0.006762199569493532, + 0.0046449871733784676, + 0.007572017144411802, + 0.008545181713998318, + 0.00896832812577486, + 0.009330404922366142, + 0.00668756477534771, + 0.007157089188694954, + 0.011996803805232048, + 0.004509045742452145, + 0.008866827003657818, + 0.007622055243700743, + 0.008804872632026672, + 0.00807703472673893, + 0.006768469233065844, + 0.005354585591703653, + 0.011848519556224346, + 0.004791056737303734, + 0.009012498892843723, + 0.006721710320562124, + 0.011839245446026325, + 0.010049897246062756, + 0.007367818616330624, + 0.009638745337724686, + 0.04147154837846756, + 0.006550066638737917, + 0.025675948709249496, + 0.015457149595022202, + 0.012862898409366608, + 0.020263122394680977, + 0.011662195436656475, + 0.03490449860692024, + 0.059231024235486984, + 0.02079971507191658, + 0.10123947262763977, + 0.02382562682032585, + 0.026354290544986725, + 0.07717517018318176, + 0.012455413118004799, + 0.006894374266266823, + 0.008342060260474682, + 0.006312470883131027, + 0.013328581117093563, + 0.01488252729177475, + 0.009800251573324203, + 0.013641390018165112, + 0.004175450187176466, + 0.004632859956473112, + 0.005803053267300129, + 0.00789082795381546, + 0.003761673578992486, + 0.005855540744960308, + 0.010456440970301628, + 0.00378043414093554, + 0.004993311129510403, + 0.00533467810600996, + 0.004237818531692028, + 0.008662731386721134, + 0.004982931073755026, + 0.009042729623615742, + 0.013599565252661705, + 0.004196970723569393, + 0.005294176284223795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0153699005022645 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637452632188797 + ], + "opposite_logprob_tokens": [ + -3.651374578475952 + ], + "logprob_reward": -0.02637452632188797, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637452632188797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08074799925088882, + 0.15996064245700836, + 0.16457273066043854, + 0.07841546833515167, + 0.11338017880916595, + 0.05346841737627983, + 0.05338485538959503, + 0.13351871073246002, + 0.029879720881581306, + 0.1326712816953659 + ], + "credit_map": [ + 0.009992231614887714, + 0.011021118611097336, + 0.0074098422192037106, + 0.007693334482610226, + 0.017682569101452827, + 0.009366678074002266, + 0.008579949848353863, + 0.00900227576494217, + 0.01926501654088497, + 0.01973617449402809, + 0.014991463162004948, + 0.02349681220948696, + 0.025040606036782265, + 0.02506684698164463, + 0.017890987917780876, + 0.01447273138910532, + 0.0161108560860157, + 0.02157318964600563, + 0.01401386596262455, + 0.020785553380846977, + 0.025471320375800133, + 0.023356035351753235, + 0.025578534230589867, + 0.01768338307738304, + 0.007175495848059654, + 0.016731899231672287, + 0.005134893581271172, + 0.010783563368022442, + 0.009592529386281967, + 0.008928008377552032, + 0.01113917026668787, + 0.008929911069571972, + 0.010432773269712925, + 0.021339455619454384, + 0.007679932750761509, + 0.018810955807566643, + 0.010920500382781029, + 0.012763692066073418, + 0.01882268488407135, + 0.012610187754034996, + 0.004063207656145096, + 0.013721750117838383, + 0.002350914990529418, + 0.008493656292557716, + 0.006377095822244883, + 0.005743239540606737, + 0.00824873149394989, + 0.004469821695238352, + 0.00671389140188694, + 0.011784074828028679, + 0.004704966209828854, + 0.009593348950147629, + 0.0052202194929122925, + 0.004971322137862444, + 0.007850744761526585, + 0.0025462908670306206, + 0.01348954439163208, + 0.018002241849899292, + 0.011205471120774746, + 0.024669628590345383, + 0.025019224733114243, + 0.009055381640791893, + 0.025090673938393593, + 0.006986544001847506, + 0.004643471911549568, + 0.0047843800857663155, + 0.00409090518951416, + 0.0031026664655655622, + 0.005018877796828747, + 0.0039495257660746574, + 0.002317871432751417, + 0.0019720224663615227, + 0.015651995316147804, + 0.008543859235942364, + 0.01756579987704754, + 0.0094436751678586, + 0.025057096034288406, + 0.026343751698732376, + 0.011091667227447033, + 0.018973445519804955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.02637452632188797 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062049854546785355 + ], + "opposite_logprob_tokens": [ + -2.8120498657226562 + ], + "logprob_reward": -0.062049854546785355, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062049854546785355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09440731257200241, + 0.12624256312847137, + 0.09586416184902191, + 0.04196792468428612, + 0.04427196830511093, + 0.06457307934761047, + 0.0517556332051754, + 0.09032834321260452, + 0.055094022303819656, + 0.3354949951171875 + ], + "credit_map": [ + 0.010653359815478325, + 0.011468091979622841, + 0.007631618995219469, + 0.010483822785317898, + 0.018759502097964287, + 0.008936109021306038, + 0.01014904584735632, + 0.01632576249539852, + 0.015267572365701199, + 0.01699690707027912, + 0.010648565366864204, + 0.016303515061736107, + 0.015987295657396317, + 0.018154334276914597, + 0.011598775163292885, + 0.021285587921738625, + 0.010775608010590076, + 0.013935824856162071, + 0.007476680912077427, + 0.013929294422268867, + 0.010076438076794147, + 0.018887849524617195, + 0.010507100261747837, + 0.010275368578732014, + 0.00414296705275774, + 0.006716677453368902, + 0.002583729103207588, + 0.00526003260165453, + 0.005562081467360258, + 0.005933824926614761, + 0.00695847999304533, + 0.004810132551938295, + 0.003139589913189411, + 0.009625483304262161, + 0.0028962770011276007, + 0.007600287441164255, + 0.004031711723655462, + 0.006241825874894857, + 0.007305540610104799, + 0.0034312524367123842, + 0.0037944086361676455, + 0.021311236545443535, + 0.0025955194141715765, + 0.012326281517744064, + 0.008304123766720295, + 0.003841064404696226, + 0.007583889178931713, + 0.004816556815057993, + 0.0037162965163588524, + 0.006549810990691185, + 0.0033557128626853228, + 0.014452404342591763, + 0.003987225703895092, + 0.005864996463060379, + 0.01157375331968069, + 0.0022554337047040462, + 0.008344108238816261, + 0.008975611999630928, + 0.007860365323722363, + 0.01619117148220539, + 0.018909724429249763, + 0.007360826712101698, + 0.018250448629260063, + 0.00443608732894063, + 0.005484454799443483, + 0.006707162130624056, + 0.006727458443492651, + 0.0057234568521380424, + 0.009039510041475296, + 0.011743579059839249, + 0.005157708656042814, + 0.0045106965117156506, + 0.027044296264648438, + 0.020588349550962448, + 0.04803561791777611, + 0.028771309182047844, + 0.053190529346466064, + 0.07473329454660416, + 0.03050648421049118, + 0.052625104784965515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.062049854546785355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002292573451996 + ], + "opposite_logprob_tokens": [ + -2.6950228214263916 + ], + "logprob_reward": -0.07002292573451996, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002292573451996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.074379101395607, + 0.10493017733097076, + 0.11012092977762222, + 0.04959537088871002, + 0.04741518571972847, + 0.06366199254989624, + 0.07444766908884048, + 0.3170347213745117, + 0.0525510311126709, + 0.10586375743150711 + ], + "credit_map": [ + 0.00901478249579668, + 0.010860234498977661, + 0.007265214342623949, + 0.00878467969596386, + 0.009541873820126057, + 0.008581825532019138, + 0.009116544388234615, + 0.011213953606784344, + 0.011099204421043396, + 0.012767311185598373, + 0.007636251859366894, + 0.015055214986205101, + 0.010222312062978745, + 0.01837247796356678, + 0.009641510434448719, + 0.020135901868343353, + 0.011265518143773079, + 0.01370618399232626, + 0.00991994608193636, + 0.014069304801523685, + 0.019668923690915108, + 0.016489177942276, + 0.01490535493940115, + 0.010096520185470581, + 0.005589472129940987, + 0.009726149961352348, + 0.0033588323276489973, + 0.0071342685259878635, + 0.004354644566774368, + 0.0061342790722846985, + 0.00888879131525755, + 0.00440893555060029, + 0.005701894871890545, + 0.00798032246530056, + 0.0027913569938391447, + 0.005678865592926741, + 0.005426512099802494, + 0.005362893454730511, + 0.007038739509880543, + 0.007434602826833725, + 0.0045464360155165195, + 0.01276510488241911, + 0.0027106408961117268, + 0.009444880299270153, + 0.009641233831644058, + 0.005783979315310717, + 0.01064010988920927, + 0.00812961533665657, + 0.00978011917322874, + 0.010910475626587868, + 0.007337502669543028, + 0.012753095477819443, + 0.0076199304312467575, + 0.006199796684086323, + 0.014862990006804466, + 0.004983763676136732, + 0.028967663645744324, + 0.03408431261777878, + 0.05461762845516205, + 0.05363203585147858, + 0.04804161936044693, + 0.019041316583752632, + 0.06181084364652634, + 0.016839342191815376, + 0.005230036564171314, + 0.006871543359011412, + 0.006315785460174084, + 0.006187782157212496, + 0.010249732993543148, + 0.008522992953658104, + 0.0053046951070427895, + 0.0038684653118252754, + 0.014893604442477226, + 0.007319411262869835, + 0.013935647904872894, + 0.008268901146948338, + 0.018027745187282562, + 0.017306126654148102, + 0.009481566958129406, + 0.01663076877593994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.07002292573451996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779873304069042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779873304069042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779873304069042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06262791901826859, + 0.08177018910646439, + 0.06843488663434982, + 0.08149734884500504, + 0.10110604763031006, + 0.10373969376087189, + 0.22725032269954681, + 0.1862352192401886, + 0.03552905470132828, + 0.051809363067150116 + ], + "credit_map": [ + 0.008652552030980587, + 0.006486536469310522, + 0.0046449583023786545, + 0.007933782413601875, + 0.012243890203535557, + 0.008511235937476158, + 0.0057062022387981415, + 0.008448761887848377, + 0.009141741320490837, + 0.006243385374546051, + 0.005385125521570444, + 0.00980570912361145, + 0.011576219461858273, + 0.012812795117497444, + 0.010107921436429024, + 0.016697296872735023, + 0.0055765677243471146, + 0.007224327884614468, + 0.005693675018846989, + 0.008913081139326096, + 0.009997548535466194, + 0.010374407283961773, + 0.012629682198166847, + 0.00802559219300747, + 0.010679759085178375, + 0.00945761613547802, + 0.005124193616211414, + 0.009880108758807182, + 0.0086248479783535, + 0.014970801770687103, + 0.010327747091650963, + 0.012432271614670753, + 0.0064764018170535564, + 0.01564801298081875, + 0.006281101144850254, + 0.014267181046307087, + 0.009172013960778713, + 0.024076595902442932, + 0.015558586455881596, + 0.009626153856515884, + 0.007224486209452152, + 0.02689526230096817, + 0.005088812671601772, + 0.017091576009988785, + 0.01374310627579689, + 0.012368391267955303, + 0.013659896329045296, + 0.007668159436434507, + 0.021992234513163567, + 0.031788188964128494, + 0.013448282144963741, + 0.06367090344429016, + 0.018833352252840996, + 0.018725913017988205, + 0.050271179527044296, + 0.008520260453224182, + 0.01652156561613083, + 0.01894601806998253, + 0.014370797201991081, + 0.02783733233809471, + 0.03485302999615669, + 0.02924645133316517, + 0.03166034817695618, + 0.012799689546227455, + 0.005037106107920408, + 0.00444477004930377, + 0.005942742805927992, + 0.004954235162585974, + 0.0038769168313592672, + 0.004687813576310873, + 0.0024556093849241734, + 0.004129858687520027, + 0.0048261587508022785, + 0.005324622616171837, + 0.006497863680124283, + 0.0047859144397079945, + 0.007261114194989204, + 0.009700588881969452, + 0.005111675709486008, + 0.00830142293125391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.009779873304069042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, ultra-high-definition, vertical medium shot captures a dynamic ink painting at eye level. Against a backdrop of Chinese landscapes that seem to come alive and flow gently across rice paper, a woman with a delicate face, long black hair loosely tied by an ancient jade hairpin, gracefully leans forward. She wears a flowing, pale-blue silk Hanfu dress, its wide sleeves gently falling with her movements. The lapels are embroidered with delicate, pale-ink bamboo leaf patterns. Her movements are focused and deliberate. A slender hand reaches down, carefully grasping the handle of a modern canvas diaper bag from the ancient bluestone slabs, weathered by time and dotted with moss, and then slowly lifts it. The off-white canvas bag stands in stark yet harmonious contrast to its classical surroundings. A vibrant, traditional auspicious cloud pattern is clearly embroidered on the front in delicate light-gold thread. The light is as soft and even as if filtering through thin rice paper. Without a clear source, it captures every detail, from the sheen of the Hanfu silk to the texture of the canvas to the woman's serene expression. As she rises, the distant mountains, rendered with thick and thin brushstrokes, continue to swirl with mist. The entire scene, nestled within the black, white, and gray of ink wash painting, is infused with vitality by a touch of pale blue on the sleeve and a touch of light gold on the embroidery. A unique atmosphere, a timeless, tranquil poetry, a blend of classical aesthetics and modern pragmatism, permeates the space.", + "seed": 13339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_12/index_7091_seed_13339_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0362858897075057, + "vlm_logprob": -0.0362858897075057, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.0362858897075057, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009896144270896912 + ], + "opposite_logprob_tokens": [ + -5.259896278381348 + ], + "logprob_reward": -0.009896144270896912, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009896144270896912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1331024169921875, + 0.1214117556810379, + 0.07193344831466675, + 0.05567793920636177, + 0.12946352362632751, + 0.1992185115814209, + 0.09904155135154724, + 0.04821806401014328, + 0.08739502727985382, + 0.05453769862651825 + ], + "credit_map": [ + 0.017058666795492172, + 0.015875093638896942, + 0.017173446714878082, + 0.026171505451202393, + 0.014399802312254906, + 0.018861694261431694, + 0.008168956264853477, + 0.015393270179629326, + 0.022743556648492813, + 0.029887013137340546, + 0.010878243483603, + 0.016526876017451286, + 0.009118318557739258, + 0.008755172602832317, + 0.013084713369607925, + 0.01041786652058363, + 0.007046905346214771, + 0.008309458382427692, + 0.009352371096611023, + 0.012516608461737633, + 0.007931076921522617, + 0.006542786955833435, + 0.011426466517150402, + 0.00880778394639492, + 0.004981592297554016, + 0.010393032804131508, + 0.006731838919222355, + 0.007013687398284674, + 0.00578328687697649, + 0.00779923377558589, + 0.006222744006663561, + 0.006752531044185162, + 0.009675104171037674, + 0.025745023041963577, + 0.010223784483969212, + 0.011244202964007854, + 0.0326089933514595, + 0.01659712754189968, + 0.013019961304962635, + 0.010349340736865997, + 0.02071237564086914, + 0.029728127643465996, + 0.01808352954685688, + 0.02902214415371418, + 0.030798939988017082, + 0.02123517356812954, + 0.028484856709837914, + 0.021153375506401062, + 0.00891047716140747, + 0.021708311513066292, + 0.009752740152180195, + 0.011204897426068783, + 0.009863704442977905, + 0.011544974520802498, + 0.010384977795183659, + 0.01567148044705391, + 0.005914973095059395, + 0.0053267464973032475, + 0.004227326717227697, + 0.008270931430161, + 0.005326171405613422, + 0.00659887446090579, + 0.006632949225604534, + 0.005920094437897205, + 0.03125881776213646, + 0.029050329700112343, + 0.006888296455144882, + 0.007721727713942528, + 0.0032490163575857878, + 0.0029190327040851116, + 0.0034392066299915314, + 0.0028686118312180042, + 0.007559864316135645, + 0.010577084496617317, + 0.0073518818244338036, + 0.0054870243184268475, + 0.007609157357364893, + 0.005772632081061602, + 0.004621266387403011, + 0.005558792036026716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.009896144270896912 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060578018426895 + ], + "opposite_logprob_tokens": [ + -3.8956058025360107 + ], + "logprob_reward": -0.02060578018426895, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060578018426895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17252683639526367, + 0.056320395320653915, + 0.0993783250451088, + 0.03301618993282318, + 0.07014953345060349, + 0.1608182191848755, + 0.12076166272163391, + 0.11302373558282852, + 0.09142695367336273, + 0.08257818967103958 + ], + "credit_map": [ + 0.02904394082725048, + 0.019739018753170967, + 0.01974131353199482, + 0.041897695511579514, + 0.012670185416936874, + 0.015442242845892906, + 0.015365652740001678, + 0.018626796081662178, + 0.009169350378215313, + 0.011197521351277828, + 0.0078255170956254, + 0.0071990410797297955, + 0.004284309688955545, + 0.0044568246230483055, + 0.006489566992968321, + 0.005698265973478556, + 0.009019671939313412, + 0.010608096607029438, + 0.01103898510336876, + 0.013170186430215836, + 0.015869127586483955, + 0.008556092157959938, + 0.020623501390218735, + 0.010492665693163872, + 0.0029922707471996546, + 0.003993050195276737, + 0.00632937066257, + 0.005735780578106642, + 0.0033997506834566593, + 0.004810421261936426, + 0.002997891278937459, + 0.0027576524298638105, + 0.005514801945537329, + 0.009685087017714977, + 0.005940335802733898, + 0.006654578261077404, + 0.016293954104185104, + 0.008375531993806362, + 0.009314228780567646, + 0.008371018804609776, + 0.014794879592955112, + 0.02972087264060974, + 0.0164934191852808, + 0.020987290889024734, + 0.028555382043123245, + 0.016927756369113922, + 0.017775006592273712, + 0.015563607215881348, + 0.010787280276417732, + 0.022589635103940964, + 0.013610619120299816, + 0.018065961077809334, + 0.010416924953460693, + 0.011267615482211113, + 0.01226161140948534, + 0.021762017160654068, + 0.019342200830578804, + 0.01576172560453415, + 0.012318775989115238, + 0.022707967087626457, + 0.009378974325954914, + 0.0087672658264637, + 0.011849394999444485, + 0.012897430919110775, + 0.022502487525343895, + 0.022089824080467224, + 0.008341762237250805, + 0.010178107768297195, + 0.0058193085715174675, + 0.008344583213329315, + 0.0072910902090370655, + 0.006859788671135902, + 0.010450486093759537, + 0.013279213570058346, + 0.010869290679693222, + 0.011698318645358086, + 0.010420472361147404, + 0.009787633083760738, + 0.007870350033044815, + 0.008202426135540009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.02060578018426895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.100298672914505 + ], + "opposite_logprob_tokens": [ + -2.3502986431121826 + ], + "logprob_reward": -0.100298672914505, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.100298672914505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20102614164352417, + 0.09598518908023834, + 0.058585260063409805, + 0.07385966181755066, + 0.0639825165271759, + 0.13375598192214966, + 0.2163240909576416, + 0.052817292511463165, + 0.046698566526174545, + 0.05696532875299454 + ], + "credit_map": [ + 0.031227990984916687, + 0.018089892342686653, + 0.024768361821770668, + 0.04016009345650673, + 0.019939279183745384, + 0.02228507027029991, + 0.014759844169020653, + 0.029795603826642036, + 0.018006879836320877, + 0.019245747476816177, + 0.01191678550094366, + 0.011547950096428394, + 0.006323777139186859, + 0.008110037073493004, + 0.010194201022386551, + 0.010639806278049946, + 0.005166732240468264, + 0.0065623559057712555, + 0.008122505620121956, + 0.008204611949622631, + 0.0072171809151768684, + 0.006936703342944384, + 0.009153937920928001, + 0.00722122797742486, + 0.006800412200391293, + 0.006659441627562046, + 0.011477206833660603, + 0.009532303549349308, + 0.0077120158821344376, + 0.013055642135441303, + 0.009636354632675648, + 0.008986284956336021, + 0.006220112554728985, + 0.016555611044168472, + 0.00403641676530242, + 0.005788787268102169, + 0.010853170417249203, + 0.007722668349742889, + 0.005434619728475809, + 0.007371127139776945, + 0.012022273615002632, + 0.021733863279223442, + 0.016937004402279854, + 0.016376283019781113, + 0.022347263991832733, + 0.01679561287164688, + 0.015583263710141182, + 0.011960404925048351, + 0.026804067194461823, + 0.06329090148210526, + 0.017194079235196114, + 0.021001100540161133, + 0.020578281953930855, + 0.028507141396403313, + 0.015333166345953941, + 0.023615345358848572, + 0.00708179222419858, + 0.00644067395478487, + 0.005937620997428894, + 0.010243195109069347, + 0.005210711620748043, + 0.004687131382524967, + 0.005321534816175699, + 0.007894632406532764, + 0.014936652034521103, + 0.013934365473687649, + 0.0037532930728048086, + 0.004039902240037918, + 0.0017795288003981113, + 0.0026738005690276623, + 0.0018045431934297085, + 0.0037764834705740213, + 0.008696956560015678, + 0.012180342338979244, + 0.00800844095647335, + 0.00740564102306962, + 0.004953269846737385, + 0.0045239911414682865, + 0.005211345851421356, + 0.005985340569168329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.100298672914505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809420391917229 + ], + "opposite_logprob_tokens": [ + -3.2880942821502686 + ], + "logprob_reward": -0.03809420391917229, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809420391917229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047165870666503906, + 0.08463230729103088, + 0.04854481294751167, + 0.057358723133802414, + 0.18674157559871674, + 0.04206383228302002, + 0.40607210993766785, + 0.08163346350193024, + 0.021817874163389206, + 0.023969458416104317 + ], + "credit_map": [ + 0.008054081350564957, + 0.0049622091464698315, + 0.005697950720787048, + 0.010611051693558693, + 0.0036278152838349342, + 0.008068633265793324, + 0.00326981907710433, + 0.0028743133880198, + 0.015589691698551178, + 0.021510234102606773, + 0.008177672512829304, + 0.011292084120213985, + 0.006027834955602884, + 0.006660367362201214, + 0.008304129354655743, + 0.00707029877230525, + 0.004319114610552788, + 0.005453576799482107, + 0.006747849751263857, + 0.0065307533368468285, + 0.00810908991843462, + 0.006480695679783821, + 0.006866496056318283, + 0.004037237726151943, + 0.0038809620309621096, + 0.006688242312520742, + 0.008769663982093334, + 0.009707506746053696, + 0.0061124167405068874, + 0.00960402563214302, + 0.005946946796029806, + 0.0066489591263234615, + 0.01583113521337509, + 0.02819439209997654, + 0.014450783841311932, + 0.015510845929384232, + 0.05058339238166809, + 0.019315635785460472, + 0.023131271824240685, + 0.01972411572933197, + 0.00313747045584023, + 0.005412379279732704, + 0.008071144111454487, + 0.005579014774411917, + 0.006434032693505287, + 0.0035473546013236046, + 0.004888947121798992, + 0.004993490409106016, + 0.03005805052816868, + 0.11437052488327026, + 0.02625122480094433, + 0.03430314362049103, + 0.040869325399398804, + 0.03498346358537674, + 0.035383641719818115, + 0.0898527279496193, + 0.010347696952521801, + 0.010958430357277393, + 0.011112945154309273, + 0.018048733472824097, + 0.006864021997898817, + 0.006255779415369034, + 0.009637241251766682, + 0.008408604189753532, + 0.003548087552189827, + 0.0033945043105632067, + 0.005025309510529041, + 0.002907747169956565, + 0.001240025507286191, + 0.0018381266854703426, + 0.0020342327188700438, + 0.0018298417562618852, + 0.003902912139892578, + 0.003478762460872531, + 0.002863396191969514, + 0.003568465355783701, + 0.0017578080296516418, + 0.003360648639500141, + 0.0026541301049292088, + 0.002383336890488863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 1 + ], + "selected_reward": -0.03809420391917229 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253464724868536 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.01253464724868536, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253464724868536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10548203438520432, + 0.12707924842834473, + 0.0653086006641388, + 0.03864927217364311, + 0.1269167810678482, + 0.22266626358032227, + 0.10113967955112457, + 0.04893269017338753, + 0.077989362180233, + 0.08583610504865646 + ], + "credit_map": [ + 0.015780113637447357, + 0.01507286075502634, + 0.011193761602044106, + 0.02110437862575054, + 0.012722383253276348, + 0.012619414366781712, + 0.009280888363718987, + 0.007708239369094372, + 0.023370105773210526, + 0.026132460683584213, + 0.011891886591911316, + 0.017978129908442497, + 0.010246661491692066, + 0.010540653020143509, + 0.014761350117623806, + 0.012158006429672241, + 0.005996836349368095, + 0.006485846359282732, + 0.01115957647562027, + 0.00847653765231371, + 0.008288497105240822, + 0.0054340907372534275, + 0.011394941247999668, + 0.008072275668382645, + 0.002515339758247137, + 0.004278915002942085, + 0.006047858390957117, + 0.004783350042998791, + 0.004525822587311268, + 0.007043210789561272, + 0.005015432834625244, + 0.004439341835677624, + 0.010689374059438705, + 0.0159233957529068, + 0.010483520105481148, + 0.011193321086466312, + 0.031336791813373566, + 0.015923427417874336, + 0.015374556183815002, + 0.015992403030395508, + 0.021997246891260147, + 0.0393323078751564, + 0.020272880792617798, + 0.028696514666080475, + 0.03946169093251228, + 0.023526495322585106, + 0.02687564492225647, + 0.02250346913933754, + 0.007476635277271271, + 0.028372684493660927, + 0.006157307419925928, + 0.006012375466525555, + 0.010813881643116474, + 0.009887377731502056, + 0.009095760062336922, + 0.02332366071641445, + 0.00650628749281168, + 0.005494341254234314, + 0.005204350221902132, + 0.008805003017187119, + 0.004998169839382172, + 0.004881119821220636, + 0.005506148096174002, + 0.007537274155765772, + 0.024961596354842186, + 0.023031868040561676, + 0.008008585311472416, + 0.007950466126203537, + 0.003366556717082858, + 0.0029313007835298777, + 0.003915790468454361, + 0.0038231872022151947, + 0.010065275244414806, + 0.014590288512408733, + 0.011901204474270344, + 0.012626929208636284, + 0.00837448425590992, + 0.008877446874976158, + 0.009260280057787895, + 0.010140199214220047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.01253464724868536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, ultra-high-definition, vertical medium shot captures a dynamic ink painting at eye level. Against a backdrop of Chinese landscapes that seem to come alive and flow gently across rice paper, a woman with a delicate face, long black hair loosely tied by an ancient jade hairpin, gracefully leans forward. She wears a flowing, pale-blue silk Hanfu dress, its wide sleeves gently falling with her movements. The lapels are embroidered with delicate, pale-ink bamboo leaf patterns. Her movements are focused and deliberate. A slender hand reaches down, carefully grasping the handle of a modern canvas diaper bag from the ancient bluestone slabs, weathered by time and dotted with moss, and then slowly lifts it. The off-white canvas bag stands in stark yet harmonious contrast to its classical surroundings. A vibrant, traditional auspicious cloud pattern is clearly embroidered on the front in delicate light-gold thread. The light is as soft and even as if filtering through thin rice paper. Without a clear source, it captures every detail, from the sheen of the Hanfu silk to the texture of the canvas to the woman's serene expression. As she rises, the distant mountains, rendered with thick and thin brushstrokes, continue to swirl with mist. The entire scene, nestled within the black, white, and gray of ink wash painting, is infused with vitality by a touch of pale blue on the sleeve and a touch of light gold on the embroidery. A unique atmosphere, a timeless, tranquil poetry, a blend of classical aesthetics and modern pragmatism, permeates the space.", + "seed": 13340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_12/index_7091_seed_13340_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02672966253012419, + "vlm_logprob": -0.02672966253012419, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02672966253012419, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009351853281259537 + ], + "opposite_logprob_tokens": [ + -5.50935173034668 + ], + "logprob_reward": -0.009351853281259537, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.009351853281259537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14696651697158813, + 0.1040959432721138, + 0.02913804166018963, + 0.16461174190044403, + 0.062457527965307236, + 0.09592381119728088, + 0.05832989886403084, + 0.05426013842225075, + 0.17443537712097168, + 0.10978108644485474 + ], + "credit_map": [ + 0.015627916902303696, + 0.013932301662862301, + 0.030025532469153404, + 0.02886996790766716, + 0.012835894711315632, + 0.024881016463041306, + 0.012225215323269367, + 0.008568665012717247, + 0.015187217853963375, + 0.009227234870195389, + 0.02680671028792858, + 0.02272537350654602, + 0.006905415561050177, + 0.007032798137515783, + 0.008079128339886665, + 0.008132056333124638, + 0.007609486114233732, + 0.0030894153751432896, + 0.004517035558819771, + 0.0053206393495202065, + 0.0021627950482070446, + 0.002049220260232687, + 0.0023712257388979197, + 0.0020182221196591854, + 0.036787353456020355, + 0.01567830890417099, + 0.009956825524568558, + 0.06339744478464127, + 0.008351008407771587, + 0.007230695802718401, + 0.01698148064315319, + 0.006228616461157799, + 0.008751020766794682, + 0.008640240877866745, + 0.0068838633596897125, + 0.008918287232518196, + 0.008004340343177319, + 0.007439610548317432, + 0.009134316816926003, + 0.004685840103775263, + 0.010804145596921444, + 0.016203904524445534, + 0.005347676575183868, + 0.007432337384670973, + 0.015402770601212978, + 0.00965520367026329, + 0.021057121455669403, + 0.01002063974738121, + 0.005029946565628052, + 0.00936646107584238, + 0.004032584838569164, + 0.004128046799451113, + 0.008883065544068813, + 0.007028297521173954, + 0.013585176318883896, + 0.006276312284171581, + 0.00738681573420763, + 0.00752602145075798, + 0.004482078831642866, + 0.004211731255054474, + 0.007078070659190416, + 0.009987708181142807, + 0.005765717010945082, + 0.007821991108357906, + 0.024399271234869957, + 0.026125481352210045, + 0.021312501281499863, + 0.024602344259619713, + 0.015236265026032925, + 0.02092227153480053, + 0.02183576487004757, + 0.02000146172940731, + 0.02236657589673996, + 0.014189297333359718, + 0.017249850556254387, + 0.014538072980940342, + 0.013053168542683125, + 0.006964069791138172, + 0.013773295097053051, + 0.0076467557810246944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.009351853281259537 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023309968411922455 + ], + "opposite_logprob_tokens": [ + -3.7733099460601807 + ], + "logprob_reward": -0.023309968411922455, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023309968411922455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08913250267505646, + 0.11141123622655869, + 0.07456658035516739, + 0.0681454986333847, + 0.057480331510305405, + 0.048103537410497665, + 0.05736494064331055, + 0.10597007721662521, + 0.2150125801563263, + 0.17281273007392883 + ], + "credit_map": [ + 0.007888509891927242, + 0.009486694820225239, + 0.014895398169755936, + 0.015799693763256073, + 0.011193132027983665, + 0.013885844498872757, + 0.008928644470870495, + 0.007054587826132774, + 0.014248829334974289, + 0.006898685358464718, + 0.02778198942542076, + 0.03494453802704811, + 0.005605960264801979, + 0.007783518172800541, + 0.008925018832087517, + 0.00522269681096077, + 0.009592216461896896, + 0.007618797477334738, + 0.006018369924277067, + 0.009828112088143826, + 0.009378256276249886, + 0.013586333952844143, + 0.010132625699043274, + 0.008411873131990433, + 0.018577927723526955, + 0.005961470305919647, + 0.003870188258588314, + 0.024172229692339897, + 0.0031793604139238596, + 0.00289398361928761, + 0.006982101127505302, + 0.002508232370018959, + 0.004230170510709286, + 0.0074364664033055305, + 0.004037031903862953, + 0.005876454990357161, + 0.013536510057747364, + 0.005873731337487698, + 0.009604373015463352, + 0.006885591894388199, + 0.00922329444438219, + 0.007343978155404329, + 0.0038658061530441046, + 0.0033398049417883158, + 0.006011348683387041, + 0.005273444578051567, + 0.008084835484623909, + 0.004961027763783932, + 0.0032479383517056704, + 0.009440546855330467, + 0.002590680494904518, + 0.0024273584131151438, + 0.01170036569237709, + 0.00865460466593504, + 0.012744981795549393, + 0.006558459717780352, + 0.011887920089066029, + 0.015593163669109344, + 0.011547454632818699, + 0.00791082251816988, + 0.01769944466650486, + 0.01229108963161707, + 0.01608416624367237, + 0.012956012040376663, + 0.022868869826197624, + 0.036948539316654205, + 0.019754832610487938, + 0.033622149378061295, + 0.023159749805927277, + 0.020367007702589035, + 0.03586670383810997, + 0.022424716502428055, + 0.01704353094100952, + 0.020616674795746803, + 0.021036000922322273, + 0.021595081314444542, + 0.022008201107382774, + 0.02321997471153736, + 0.03124307282269001, + 0.016050197184085846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.023309968411922455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203921139240265 + ], + "opposite_logprob_tokens": [ + -2.8120391368865967 + ], + "logprob_reward": -0.06203921139240265, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203921139240265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09157028794288635, + 0.05608247593045235, + 0.030703512951731682, + 0.053957559168338776, + 0.05492505803704262, + 0.05821976438164711, + 0.1283653974533081, + 0.2276614010334015, + 0.170598104596138, + 0.12791649997234344 + ], + "credit_map": [ + 0.010815659537911415, + 0.007680251728743315, + 0.019284382462501526, + 0.019931504502892494, + 0.006975013762712479, + 0.015102237462997437, + 0.006120902020484209, + 0.005660331342369318, + 0.007631008513271809, + 0.004504813347011805, + 0.013630209490656853, + 0.014396014623343945, + 0.0032800212502479553, + 0.004460844676941633, + 0.005010849330574274, + 0.003168718423694372, + 0.0064932783134281635, + 0.0035002059303224087, + 0.005254132207483053, + 0.005425481125712395, + 0.0020517578814178705, + 0.002897035563364625, + 0.0026933117769658566, + 0.0023883075919002295, + 0.009573079645633698, + 0.006325528956949711, + 0.003930060658603907, + 0.020582368597388268, + 0.002744989236816764, + 0.0020098944660276175, + 0.006829465739428997, + 0.0019621732644736767, + 0.0076344809494912624, + 0.007362157106399536, + 0.008581135421991348, + 0.006940115708857775, + 0.0070828828029334545, + 0.005410868674516678, + 0.005627723410725594, + 0.006285690236836672, + 0.008222050033509731, + 0.008059813641011715, + 0.0071974643506109715, + 0.005477400962263346, + 0.00920685101300478, + 0.005344589706510305, + 0.008641181513667107, + 0.006070411764085293, + 0.009924495592713356, + 0.017910635098814964, + 0.006677397061139345, + 0.009777156636118889, + 0.020734308287501335, + 0.020326290279626846, + 0.023505574092268944, + 0.019509531557559967, + 0.02375052683055401, + 0.030457250773906708, + 0.029501376673579216, + 0.0196318831294775, + 0.035006988793611526, + 0.03045748360455036, + 0.030836010351777077, + 0.02801988460123539, + 0.020764170214533806, + 0.0229155495762825, + 0.01852753758430481, + 0.033912863582372665, + 0.02407005988061428, + 0.02096850797533989, + 0.016348127275705338, + 0.013091298751533031, + 0.018795963376760483, + 0.015893083065748215, + 0.015438525006175041, + 0.01863519847393036, + 0.01925673522055149, + 0.00969846174120903, + 0.017768604680895805, + 0.01242993026971817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.06203921139240265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369775414466858 + ], + "opposite_logprob_tokens": [ + -3.4086978435516357 + ], + "logprob_reward": -0.03369775414466858, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369775414466858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09332087635993958, + 0.054338086396455765, + 0.039658475667238235, + 0.03551134839653969, + 0.02862834557890892, + 0.05826554819941521, + 0.06092897802591324, + 0.09652478992938995, + 0.3565693497657776, + 0.17625418305397034 + ], + "credit_map": [ + 0.011777592822909355, + 0.007609360385686159, + 0.018900780007243156, + 0.020220205187797546, + 0.008045690134167671, + 0.014527923427522182, + 0.006939012091606855, + 0.005300309509038925, + 0.008106679655611515, + 0.004179332405328751, + 0.012526252306997776, + 0.014942592009902, + 0.0031695484649389982, + 0.003697272390127182, + 0.004584384150803089, + 0.0031320236157625914, + 0.010430452413856983, + 0.0037204946856945753, + 0.007916145958006382, + 0.007977130822837353, + 0.002757077105343342, + 0.002068864880129695, + 0.002860890468582511, + 0.0019274207297712564, + 0.004054373130202293, + 0.005961536429822445, + 0.003908049780875444, + 0.011134536936879158, + 0.003288462059572339, + 0.0013674653600901365, + 0.004278008826076984, + 0.0015189154073596, + 0.005396946333348751, + 0.0037044761702418327, + 0.004385658539831638, + 0.004168021958321333, + 0.003215960692614317, + 0.0017793121514841914, + 0.0035773615818470716, + 0.002400607569143176, + 0.010382622480392456, + 0.008246595039963722, + 0.004774356260895729, + 0.00454045319929719, + 0.009670545347034931, + 0.004685893189162016, + 0.009992070496082306, + 0.005973012186586857, + 0.0049380091950297356, + 0.009146698750555515, + 0.0024589376989752054, + 0.0027932696975767612, + 0.01000785268843174, + 0.010825521312654018, + 0.012731313705444336, + 0.008027378469705582, + 0.008798868395388126, + 0.014095342718064785, + 0.01151270605623722, + 0.010341067798435688, + 0.015910102054476738, + 0.011476346291601658, + 0.012756834737956524, + 0.01163352932780981, + 0.03582323342561722, + 0.052195630967617035, + 0.032309986650943756, + 0.04462139680981636, + 0.034857962280511856, + 0.04400847479701042, + 0.05723535642027855, + 0.05551731213927269, + 0.022065062075853348, + 0.017871305346488953, + 0.022402910515666008, + 0.01843426190316677, + 0.03355303406715393, + 0.022660313174128532, + 0.028758792206645012, + 0.010508508421480656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.03369775414466858 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene presented in a vertical medium shot with a cinematic ink painting style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a pale-blue silk Hanfu dress with bamboo leaf embroidery visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman gracefully lean forward to grasp a modern canvas diaper bag from ancient bluestone slabs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the diaper bag off-white with a vibrant traditional auspicious cloud pattern embroidered in light-gold thread?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a flowing Chinese landscape on rice paper with distant mountains rendered in thick and thin brushstrokes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249525420367718 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249525420367718, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249525420367718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13007253408432007, + 0.10926461964845657, + 0.09657824039459229, + 0.03774697706103325, + 0.06535273790359497, + 0.0933748409152031, + 0.055811233818531036, + 0.08687157183885574, + 0.1821662187576294, + 0.1427610218524933 + ], + "credit_map": [ + 0.020781895145773888, + 0.012095486745238304, + 0.023286642506718636, + 0.024722132831811905, + 0.009784365072846413, + 0.018282534554600716, + 0.011534901335835457, + 0.009584572166204453, + 0.01639050617814064, + 0.00857793353497982, + 0.02692272700369358, + 0.022367428988218307, + 0.007165143731981516, + 0.009044698439538479, + 0.011819170787930489, + 0.006977007258683443, + 0.024544553831219673, + 0.0097314128652215, + 0.011077897623181343, + 0.015968672931194305, + 0.007095944136381149, + 0.010880297981202602, + 0.011560247279703617, + 0.00571921793743968, + 0.004180910997092724, + 0.006014473736286163, + 0.004841253627091646, + 0.0049724518321454525, + 0.00583268515765667, + 0.0035959607921540737, + 0.0048753200098872185, + 0.0034339220728725195, + 0.00935092568397522, + 0.008117577992379665, + 0.011675327084958553, + 0.01114756241440773, + 0.00859594251960516, + 0.005366807337850332, + 0.0068601686507463455, + 0.004238427616655827, + 0.010314876213669777, + 0.014498881995677948, + 0.009191013872623444, + 0.006859108805656433, + 0.009965694509446621, + 0.010210361331701279, + 0.020992135629057884, + 0.011342772282660007, + 0.004487545695155859, + 0.008529635146260262, + 0.004437742754817009, + 0.005048729479312897, + 0.008097036741673946, + 0.006006167270243168, + 0.01199903804808855, + 0.007205337751656771, + 0.008326089009642601, + 0.013410280458629131, + 0.006927168928086758, + 0.008096957579255104, + 0.021220747381448746, + 0.009267437271773815, + 0.012727566063404083, + 0.006895327474921942, + 0.021655414253473282, + 0.02761748433113098, + 0.019252516329288483, + 0.03223225101828575, + 0.018498964607715607, + 0.017316849902272224, + 0.02487918548285961, + 0.020713569596409798, + 0.018804440274834633, + 0.017783716320991516, + 0.024939198046922684, + 0.020738545805215836, + 0.017323022708296776, + 0.010466979816555977, + 0.023188011720776558, + 0.009517109952867031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.005249525420367718 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_13/index_tensor([8480, 8480])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_13/index_tensor([8480, 8480])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..43e14762b87ad7742a6b273df19cb1d48aef9412 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_13/index_tensor([8480, 8480])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, taken from a slightly low angle, captures a scene immersed in the surreal Makoto Shinkai aesthetic, the air itself brimming with emotion and light. The room is bathed in intense, golden late-afternoon sunlight, which filters through a large window draped with sheer white curtains, casting long, soft shadows and illuminating millions of dust particles dancing across the space in the visible twilight. On a light-colored wooden table lies a pristine stack of five vibrantly colored light novels, their glossy covers adorned with vibrant, saturated, anime-style character art. As if weightless, the entire volume rises smoothly and deliberately from the table's surface, a faint, ethereal aura shimmering beneath them. Once lifted into the air, the bookshelf gracefully parted, and each book began to gently rotate in the center of the room, a slow, enchanting ballet. The pages of the topmost book, with its vibrant pink and blue cover, began to flutter open, as if blown by a silent, magical breeze, revealing intricate illustrations from which emanated a brilliant, warm light, softly pulsing and casting a dynamic presence that shifted highlights throughout the room, creating a tranquil, breathtaking atmosphere.", + "seed": 12501, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_13/index_8480_seed_12501_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009235988277941943, + "vlm_logprob": -0.009235988277941943, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.009235988277941943, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0056891427375376225 + ], + "opposite_logprob_tokens": [ + -6.0056891441345215 + ], + "logprob_reward": -0.0056891427375376225, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0056891427375376225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1110091358423233, + 0.08512164652347565, + 0.11047758907079697, + 0.2159050554037094, + 0.12178631871938705, + 0.10325441509485245, + 0.04604542255401611, + 0.0472775474190712, + 0.07271946221590042, + 0.08640336990356445 + ], + "credit_map": [ + 0.018054481595754623, + 0.018641091883182526, + 0.010283094830811024, + 0.011959298513829708, + 0.010861493647098541, + 0.010539927519857883, + 0.014235525391995907, + 0.016434216871857643, + 0.011295881122350693, + 0.013499176129698753, + 0.009439270943403244, + 0.008101597428321838, + 0.006032009143382311, + 0.017930911853909492, + 0.011068070307374, + 0.007754726335406303, + 0.01420833170413971, + 0.03111288510262966, + 0.005995979532599449, + 0.004655085504055023, + 0.009905118495225906, + 0.03720627352595329, + 0.0033965716138482094, + 0.003997343592345715, + 0.02886977791786194, + 0.026155469939112663, + 0.01858595758676529, + 0.025836719200015068, + 0.02572011761367321, + 0.037514448165893555, + 0.03947020322084427, + 0.013752363622188568, + 0.013483398593962193, + 0.018487850204110146, + 0.010594693012535572, + 0.009931044653058052, + 0.028872810304164886, + 0.016713345423340797, + 0.015687059611082077, + 0.00801611877977848, + 0.012713108211755753, + 0.013896605931222439, + 0.008794495835900307, + 0.014325113967061043, + 0.01935390569269657, + 0.011960760690271854, + 0.015252910554409027, + 0.00695751653984189, + 0.005833832081407309, + 0.008518291637301445, + 0.004120273049920797, + 0.003961689304560423, + 0.008150560781359673, + 0.006398525554686785, + 0.005829731933772564, + 0.00323251960799098, + 0.004287548828870058, + 0.009683062322437763, + 0.004202778451144695, + 0.006501653231680393, + 0.008609041571617126, + 0.00445563206449151, + 0.00567574892193079, + 0.003862083423882723, + 0.007412125822156668, + 0.015642140060663223, + 0.006038253661245108, + 0.005902386270463467, + 0.012604859657585621, + 0.006865795236080885, + 0.009589429944753647, + 0.008664470165967941, + 0.00874209962785244, + 0.018512984737753868, + 0.007903381250798702, + 0.008982683531939983, + 0.017500776797533035, + 0.008952664211392403, + 0.008016673848032951, + 0.007792109157890081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.0056891427375376225 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028364923782646656 + ], + "opposite_logprob_tokens": [ + -5.87783670425415 + ], + "logprob_reward": -0.0028364923782646656, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028364923782646656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30498090386390686, + 0.06037062034010887, + 0.03915294259786606, + 0.0548451691865921, + 0.05013260617852211, + 0.10946815460920334, + 0.09615132957696915, + 0.043314069509506226, + 0.06764044612646103, + 0.17394380271434784 + ], + "credit_map": [ + 0.048491161316633224, + 0.029865095391869545, + 0.04336537793278694, + 0.033183928579092026, + 0.01836903765797615, + 0.049894750118255615, + 0.03654770180583, + 0.04526384547352791, + 0.010390917770564556, + 0.007383513264358044, + 0.008426109328866005, + 0.00620223069563508, + 0.003999439999461174, + 0.006698958575725555, + 0.007219307590276003, + 0.01005014032125473, + 0.005860397592186928, + 0.00576125318184495, + 0.0041905660182237625, + 0.004202370066195726, + 0.005017284769564867, + 0.0060034701600670815, + 0.0037430322263389826, + 0.004374567419290543, + 0.007161838933825493, + 0.009340384975075722, + 0.004933066200464964, + 0.004507102072238922, + 0.009553293697535992, + 0.008916892111301422, + 0.006748207379132509, + 0.003684382187202573, + 0.006308469455689192, + 0.006613187957555056, + 0.005139532033354044, + 0.005538819823414087, + 0.007708944380283356, + 0.006805571261793375, + 0.007305127568542957, + 0.004712951835244894, + 0.013027729466557503, + 0.01851349137723446, + 0.010969653725624084, + 0.012099812738597393, + 0.022375544533133507, + 0.011670983396470547, + 0.014796332456171513, + 0.006014606915414333, + 0.010610049590468407, + 0.021124636754393578, + 0.010517475195229053, + 0.007559433579444885, + 0.017216462641954422, + 0.010837187059223652, + 0.010120722465217113, + 0.008165359497070312, + 0.0054626138880848885, + 0.009671760722994804, + 0.0042694429866969585, + 0.004813916515558958, + 0.00574202835559845, + 0.0029871598817408085, + 0.004564871080219746, + 0.005802276078611612, + 0.006722064223140478, + 0.017812678590416908, + 0.005909043364226818, + 0.00698766577988863, + 0.013142594136297703, + 0.0051915934309363365, + 0.00822498369961977, + 0.0036498229019343853, + 0.019915707409381866, + 0.024644766002893448, + 0.020326146855950356, + 0.01334991492331028, + 0.03793041780591011, + 0.01646181382238865, + 0.01582235097885132, + 0.025492675602436066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.0028364923782646656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215044677257538 + ], + "opposite_logprob_tokens": [ + -4.264214992523193 + ], + "logprob_reward": -0.014215044677257538, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215044677257538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08477386832237244, + 0.11724632978439331, + 0.34206560254096985, + 0.07447410374879837, + 0.02828170731663704, + 0.20203834772109985, + 0.03174411132931709, + 0.05859978497028351, + 0.030737072229385376, + 0.030039090663194656 + ], + "credit_map": [ + 0.01057075709104538, + 0.013279189355671406, + 0.008476164191961288, + 0.008539952337741852, + 0.005975143052637577, + 0.017982251942157745, + 0.011304355226457119, + 0.00864605512470007, + 0.018075555562973022, + 0.020391156896948814, + 0.012407849542796612, + 0.009993582963943481, + 0.009252204559743404, + 0.024729719385504723, + 0.012417428195476532, + 0.009978835470974445, + 0.05258100479841232, + 0.074379563331604, + 0.02063802257180214, + 0.016400348395109177, + 0.030046949163079262, + 0.11784631758928299, + 0.016319142654538155, + 0.013854239135980606, + 0.009571545757353306, + 0.012502762489020824, + 0.006951316259801388, + 0.009305162355303764, + 0.011146873235702515, + 0.011522602289915085, + 0.00804008450359106, + 0.005433756858110428, + 0.003313411958515644, + 0.004191718064248562, + 0.003853287547826767, + 0.002405239501968026, + 0.005275070201605558, + 0.003647706238552928, + 0.0037175831384956837, + 0.0018776892684400082, + 0.024508008733391762, + 0.03524108976125717, + 0.02149450220167637, + 0.02690308168530464, + 0.041864633560180664, + 0.01860993541777134, + 0.022609954699873924, + 0.010807139798998833, + 0.0039809769950807095, + 0.005786761175841093, + 0.004087541252374649, + 0.0030572714749723673, + 0.004648653324693441, + 0.0025997606571763754, + 0.004550629295408726, + 0.0030325190164148808, + 0.00699802627786994, + 0.011840859428048134, + 0.006521323695778847, + 0.005105982068926096, + 0.008355855010449886, + 0.0055143218487501144, + 0.00569869065657258, + 0.008564723655581474, + 0.0031180179212242365, + 0.008385403081774712, + 0.003196009434759617, + 0.0023530288599431515, + 0.006048819050192833, + 0.0021618884056806564, + 0.0027604682836681604, + 0.0027134381234645844, + 0.0026716056745499372, + 0.006299085449427366, + 0.0022983436938375235, + 0.0038094669580459595, + 0.00523746944963932, + 0.002305794507265091, + 0.005391769576817751, + 0.0020255553536117077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.014215044677257538 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818273961544037 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.01818273961544037, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818273961544037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16465336084365845, + 0.041150886565446854, + 0.14133425056934357, + 0.046571820974349976, + 0.037806931883096695, + 0.07120124250650406, + 0.24732908606529236, + 0.11122427880764008, + 0.07836293429136276, + 0.0603652186691761 + ], + "credit_map": [ + 0.025663021951913834, + 0.027624668553471565, + 0.013714822009205818, + 0.011301099322736263, + 0.010409773327410221, + 0.04932340979576111, + 0.012531772255897522, + 0.01408478431403637, + 0.006025272887200117, + 0.0069845509715378284, + 0.005231519695371389, + 0.0039815581403672695, + 0.00352234928868711, + 0.009365096688270569, + 0.0029273205436766148, + 0.0031132197473198175, + 0.02179563231766224, + 0.029053404927253723, + 0.008560244925320148, + 0.007886276580393314, + 0.012729393318295479, + 0.04799797758460045, + 0.007358168717473745, + 0.005953149404376745, + 0.0055220299400389194, + 0.007443918380886316, + 0.003891740459948778, + 0.007290646433830261, + 0.006062278524041176, + 0.006158782634884119, + 0.00696130795404315, + 0.003241117112338543, + 0.005286748055368662, + 0.0046014138497412205, + 0.004574230872094631, + 0.003572338493540883, + 0.005226885434240103, + 0.006830996368080378, + 0.005750478710979223, + 0.001963841263204813, + 0.009812328964471817, + 0.009309404529631138, + 0.010486651211977005, + 0.007123537361621857, + 0.010346762835979462, + 0.007810491602867842, + 0.010947292670607567, + 0.005364777520298958, + 0.030733490362763405, + 0.05239947512745857, + 0.028140626847743988, + 0.018816707655787468, + 0.04093591868877411, + 0.031650085002183914, + 0.023797383531928062, + 0.020855404436588287, + 0.012689532712101936, + 0.027369139716029167, + 0.012178044766187668, + 0.013009178452193737, + 0.016866568475961685, + 0.00991037767380476, + 0.012083364650607109, + 0.007118068635463715, + 0.006838483735918999, + 0.02274877391755581, + 0.006332884542644024, + 0.005718372296541929, + 0.014938926324248314, + 0.008333177305758, + 0.006059936713427305, + 0.007392382249236107, + 0.006353135220706463, + 0.010237662121653557, + 0.005257424432784319, + 0.006567169912159443, + 0.007940449751913548, + 0.0056881411001086235, + 0.012505444698035717, + 0.0058157965540885925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.01818273961544037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052565219812095165 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052565219812095165, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052565219812095165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0774196982383728, + 0.0735936239361763, + 0.18135827779769897, + 0.046360164880752563, + 0.027900967746973038, + 0.20553818345069885, + 0.17507480084896088, + 0.058036644011735916, + 0.06305529177188873, + 0.09166234731674194 + ], + "credit_map": [ + 0.012006806209683418, + 0.0075373100116848946, + 0.00676314253360033, + 0.009577810764312744, + 0.004589005373418331, + 0.01514450367540121, + 0.016002118587493896, + 0.005799002014100552, + 0.011780223809182644, + 0.011848325841128826, + 0.005273289047181606, + 0.004584512673318386, + 0.005522603634744883, + 0.022172633558511734, + 0.00849364884197712, + 0.003918388392776251, + 0.026935722678899765, + 0.04439135268330574, + 0.010323193855583668, + 0.006794413551688194, + 0.015067624859511852, + 0.0652947723865509, + 0.006269074510782957, + 0.006282133515924215, + 0.007968263700604439, + 0.005603103898465633, + 0.0036895889788866043, + 0.003965721465647221, + 0.004661222919821739, + 0.01232972089201212, + 0.005243123508989811, + 0.002899416256695986, + 0.003705941839143634, + 0.0033655583392828703, + 0.0024387987796217203, + 0.002503148512914777, + 0.005852086003869772, + 0.004220270086079836, + 0.0034035933203995228, + 0.002411570865660906, + 0.02559880167245865, + 0.036044489592313766, + 0.025677796453237534, + 0.02349139377474785, + 0.041928380727767944, + 0.01818985491991043, + 0.023447955027222633, + 0.011159508489072323, + 0.019522862508893013, + 0.041206344962120056, + 0.023759109899401665, + 0.014056481420993805, + 0.027754971757531166, + 0.017497647553682327, + 0.01834854483604431, + 0.012928842566907406, + 0.0068176258355379105, + 0.01487629022449255, + 0.00517669040709734, + 0.006267497316002846, + 0.009233691729605198, + 0.004858036991208792, + 0.005819736979901791, + 0.004987070802599192, + 0.005278059747070074, + 0.01831924170255661, + 0.004640111234039068, + 0.005056832451373339, + 0.011367582716047764, + 0.005734769627451897, + 0.005895215552300215, + 0.0067634740844368935, + 0.011158300563693047, + 0.018431998789310455, + 0.006785512436181307, + 0.009897933341562748, + 0.018950412049889565, + 0.00844625849276781, + 0.012356195598840714, + 0.005635733250528574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.0052565219812095165 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, taken from a slightly low angle, captures a scene immersed in the surreal Makoto Shinkai aesthetic, the air itself brimming with emotion and light. The room is bathed in intense, golden late-afternoon sunlight, which filters through a large window draped with sheer white curtains, casting long, soft shadows and illuminating millions of dust particles dancing across the space in the visible twilight. On a light-colored wooden table lies a pristine stack of five vibrantly colored light novels, their glossy covers adorned with vibrant, saturated, anime-style character art. As if weightless, the entire volume rises smoothly and deliberately from the table's surface, a faint, ethereal aura shimmering beneath them. Once lifted into the air, the bookshelf gracefully parted, and each book began to gently rotate in the center of the room, a slow, enchanting ballet. The pages of the topmost book, with its vibrant pink and blue cover, began to flutter open, as if blown by a silent, magical breeze, revealing intricate illustrations from which emanated a brilliant, warm light, softly pulsing and casting a dynamic presence that shifted highlights throughout the room, creating a tranquil, breathtaking atmosphere.", + "seed": 12502, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_13/index_8480_seed_12502_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.007015821989625692, + "vlm_logprob": -0.007015821989625692, + "vlm_yes_no_margin": 5.125, + "vlm_token_credit": -0.007015821989625692, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003999806474894285 + ], + "opposite_logprob_tokens": [ + -5.878999710083008 + ], + "logprob_reward": -0.003999806474894285, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.003999806474894285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23652517795562744, + 0.10635928064584732, + 0.1274721473455429, + 0.047248512506484985, + 0.07386928796768188, + 0.08407281339168549, + 0.0732397511601448, + 0.11419456452131271, + 0.08598548173904419, + 0.05103296786546707 + ], + "credit_map": [ + 0.027747156098484993, + 0.022612784057855606, + 0.02204015478491783, + 0.033770084381103516, + 0.023761747404932976, + 0.0431709960103035, + 0.0402802973985672, + 0.023141968995332718, + 0.011062365025281906, + 0.015482446178793907, + 0.011602737940847874, + 0.007680137176066637, + 0.00997166521847248, + 0.022061487659811974, + 0.00541155319660902, + 0.023086879402399063, + 0.013049974106252193, + 0.0277961827814579, + 0.011811165139079094, + 0.008955846540629864, + 0.01668480597436428, + 0.017205048352479935, + 0.008613541722297668, + 0.023355593904852867, + 0.0037861200980842113, + 0.007771449629217386, + 0.005084672477096319, + 0.004422865808010101, + 0.0068376325070858, + 0.006270593032240868, + 0.005644586868584156, + 0.007430592551827431, + 0.015252730809152126, + 0.01284928061068058, + 0.007150447461754084, + 0.009222062304615974, + 0.009134821593761444, + 0.007660228293389082, + 0.006467226427048445, + 0.0061324890702962875, + 0.009417475201189518, + 0.016149044036865234, + 0.010177519172430038, + 0.008058479987084866, + 0.013111069798469543, + 0.007564652245491743, + 0.009869605302810669, + 0.009724964387714863, + 0.009249885566532612, + 0.011078949086368084, + 0.009600656107068062, + 0.011567819863557816, + 0.008473637513816357, + 0.006390745285898447, + 0.006253169383853674, + 0.010624888353049755, + 0.016225218772888184, + 0.015006219036877155, + 0.01122038159519434, + 0.011784881353378296, + 0.013540643267333508, + 0.013162078335881233, + 0.011337348259985447, + 0.0219177957624197, + 0.007114286534488201, + 0.009405413642525673, + 0.0061963084153831005, + 0.014578400179743767, + 0.01516422163695097, + 0.007058413233608007, + 0.020638057962059975, + 0.00583037780597806, + 0.0063855284824967384, + 0.004314253106713295, + 0.008398539386689663, + 0.007230388466268778, + 0.003931386396288872, + 0.008407002314925194, + 0.0068746828474104404, + 0.005491189192980528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.003999806474894285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643728025257587 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.004643728025257587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643728025257587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1138693243265152, + 0.03598540276288986, + 0.03711049258708954, + 0.07171723246574402, + 0.07307787239551544, + 0.06754743307828903, + 0.08679467439651489, + 0.3663027286529541, + 0.05675014480948448, + 0.09084466844797134 + ], + "credit_map": [ + 0.021388379856944084, + 0.013744130730628967, + 0.011716223321855068, + 0.011128289625048637, + 0.011795477010309696, + 0.023777736350893974, + 0.007533860858529806, + 0.012785234488546848, + 0.004075841046869755, + 0.003277770010754466, + 0.006400653626769781, + 0.004383916966617107, + 0.003825318068265915, + 0.003821983002126217, + 0.0029643934685736895, + 0.0072355326265096664, + 0.0031333183869719505, + 0.005226436071097851, + 0.004836341366171837, + 0.004282183479517698, + 0.005953790154308081, + 0.0042305015958845615, + 0.003314447123557329, + 0.00613347627222538, + 0.006098561454564333, + 0.011391338892281055, + 0.00870392844080925, + 0.00478410255163908, + 0.013496795669198036, + 0.007893245667219162, + 0.007252905517816544, + 0.012096360325813293, + 0.01033494621515274, + 0.009398152120411396, + 0.007245547603815794, + 0.01214570365846157, + 0.0077110109850764275, + 0.00720288697630167, + 0.009359129704535007, + 0.00968050304800272, + 0.006171810906380415, + 0.009464800357818604, + 0.010977438651025295, + 0.008477205410599709, + 0.008851143531501293, + 0.006962314713746309, + 0.0068901702761650085, + 0.009752551093697548, + 0.009799056686460972, + 0.013088084757328033, + 0.011738291010260582, + 0.007857464253902435, + 0.012601172551512718, + 0.008367071859538555, + 0.008674572221934795, + 0.014668970368802547, + 0.04504900798201561, + 0.06238020211458206, + 0.03250264376401901, + 0.029033271595835686, + 0.07949016243219376, + 0.035972703248262405, + 0.03155864030122757, + 0.05031612515449524, + 0.0061056725680828094, + 0.00773289380595088, + 0.007919508032500744, + 0.008267535828053951, + 0.009134934283792973, + 0.005116121377795935, + 0.00596787640824914, + 0.006505604367703199, + 0.016647757962346077, + 0.005805197637528181, + 0.01296659279614687, + 0.007386285811662674, + 0.006670271512120962, + 0.01849723793566227, + 0.006535272113978863, + 0.01633606106042862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.004643728025257587 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648676797747612 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.008648676797747612, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648676797747612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24309247732162476, + 0.05869273096323013, + 0.06585607677698135, + 0.03960772976279259, + 0.044025860726833344, + 0.06321793794631958, + 0.06862283498048782, + 0.15803490579128265, + 0.17873826622962952, + 0.08011112362146378 + ], + "credit_map": [ + 0.01918213441967964, + 0.02098245918750763, + 0.02294236794114113, + 0.04933570697903633, + 0.018980201333761215, + 0.023460393771529198, + 0.053755950182676315, + 0.034453269094228745, + 0.00685259560123086, + 0.008413398638367653, + 0.006836569402366877, + 0.007424205541610718, + 0.006831873673945665, + 0.009213640354573727, + 0.005399142857640982, + 0.007721299771219492, + 0.007213597651571035, + 0.010828135535120964, + 0.00851225946098566, + 0.008767594583332539, + 0.007776190992444754, + 0.008827580139040947, + 0.006785199977457523, + 0.007145521696656942, + 0.004775723908096552, + 0.0062921857461333275, + 0.003506741253659129, + 0.006901903077960014, + 0.005571478977799416, + 0.004050770774483681, + 0.0043714940547943115, + 0.004137430340051651, + 0.007153068203479052, + 0.006253561936318874, + 0.0036156305577605963, + 0.004963492043316364, + 0.006540906149893999, + 0.004590962547808886, + 0.005902909208089113, + 0.005005332175642252, + 0.006222139112651348, + 0.015177411027252674, + 0.008129560388624668, + 0.006694194860756397, + 0.010174574330449104, + 0.00554623082280159, + 0.006352744996547699, + 0.004921081941574812, + 0.00867367535829544, + 0.010702556930482388, + 0.006446679588407278, + 0.006764467339962721, + 0.00892700906842947, + 0.006762141361832619, + 0.010934059508144855, + 0.009412246756255627, + 0.015939490869641304, + 0.02153843455016613, + 0.013193842023611069, + 0.01613520458340645, + 0.02477899193763733, + 0.015315881930291653, + 0.02060605213046074, + 0.030527008697390556, + 0.012747284956276417, + 0.024149304255843163, + 0.013287649489939213, + 0.0213303342461586, + 0.03194195404648781, + 0.020520273596048355, + 0.0352962464094162, + 0.019465221092104912, + 0.009359640069305897, + 0.009772741235792637, + 0.012624210678040981, + 0.01325454656034708, + 0.005241775419563055, + 0.011904158629477024, + 0.008372025564312935, + 0.00958202313631773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.008648676797747612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534882873296738 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.012534882873296738, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534882873296738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22889851033687592, + 0.028515717014670372, + 0.023836670443415642, + 0.039785172790288925, + 0.030267251655459404, + 0.10771050304174423, + 0.08235278725624084, + 0.22563473880290985, + 0.07987914234399796, + 0.1531195044517517 + ], + "credit_map": [ + 0.03465760499238968, + 0.018586402758955956, + 0.019700627774000168, + 0.04335387051105499, + 0.018354693427681923, + 0.022303912788629532, + 0.04609174281358719, + 0.025849657133221626, + 0.002626169705763459, + 0.003061617026105523, + 0.0037813640665262938, + 0.004408571403473616, + 0.0033396538347005844, + 0.0033013173379004, + 0.003212409559637308, + 0.004784613382071257, + 0.002981818513944745, + 0.004139193799346685, + 0.0024051046930253506, + 0.002609402406960726, + 0.00341674848459661, + 0.002960796235129237, + 0.0022355415858328342, + 0.00308806449174881, + 0.003193969838321209, + 0.005550172179937363, + 0.0032413280569016933, + 0.006525389384478331, + 0.004504989366978407, + 0.0042027197778224945, + 0.0072049652226269245, + 0.005361639894545078, + 0.004204263910651207, + 0.003954828716814518, + 0.0038778006564825773, + 0.00387580506503582, + 0.003576483577489853, + 0.0033329008147120476, + 0.0034032545518130064, + 0.004041914828121662, + 0.01227634958922863, + 0.01633830927312374, + 0.017741434276103973, + 0.009480963461101055, + 0.014779586344957352, + 0.01115661021322012, + 0.010883797891438007, + 0.015053457580506802, + 0.008353261277079582, + 0.013161692768335342, + 0.011093052104115486, + 0.011516851373016834, + 0.01052224263548851, + 0.006756065413355827, + 0.008743299171328545, + 0.012206322513520718, + 0.01806453801691532, + 0.04236256703734398, + 0.01422247663140297, + 0.03185069561004639, + 0.03532394766807556, + 0.022910060361027718, + 0.02887563593685627, + 0.032024819403886795, + 0.008770964108407497, + 0.011098094284534454, + 0.008016145788133144, + 0.010404089465737343, + 0.014141630381345749, + 0.007475442718714476, + 0.011399375274777412, + 0.008573401719331741, + 0.02001318521797657, + 0.015013000927865505, + 0.019227225333452225, + 0.02051440440118313, + 0.013019541278481483, + 0.026135241612792015, + 0.020998630672693253, + 0.018198257312178612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.012534882873296738 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052520157769322395 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.0052520157769322395, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052520157769322395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35057029128074646, + 0.04281855374574661, + 0.03743990883231163, + 0.02108188346028328, + 0.020300835371017456, + 0.04106359928846359, + 0.037775617092847824, + 0.17534072697162628, + 0.11662351340055466, + 0.1569850742816925 + ], + "credit_map": [ + 0.02977408654987812, + 0.02315548062324524, + 0.028169764205813408, + 0.040569599717855453, + 0.025646863505244255, + 0.09871462732553482, + 0.04208751022815704, + 0.06245235726237297, + 0.00408916175365448, + 0.003219987265765667, + 0.004973015747964382, + 0.007418869994580746, + 0.0027258938644081354, + 0.007551661692559719, + 0.00609203428030014, + 0.006747927982360125, + 0.003138917498290539, + 0.00654158229008317, + 0.0035594471264630556, + 0.005596340633928776, + 0.004459231626242399, + 0.004906161688268185, + 0.0037971080746501684, + 0.005441118963062763, + 0.0021763730328530073, + 0.003886248217895627, + 0.0024581432808190584, + 0.0022460140753537416, + 0.002919569378718734, + 0.0020237253047525883, + 0.0023210993967950344, + 0.003050710540264845, + 0.0023596822284162045, + 0.0026034831535071135, + 0.0021066281478852034, + 0.0027027386240661144, + 0.0029830497223883867, + 0.002664345083758235, + 0.0022949802223592997, + 0.0025859300512820482, + 0.004909667652100325, + 0.005465451627969742, + 0.00630510039627552, + 0.002710181986913085, + 0.005621024407446384, + 0.004579407628625631, + 0.005242746789008379, + 0.00623001903295517, + 0.003948835656046867, + 0.005754140205681324, + 0.004169769585132599, + 0.004587574861943722, + 0.00546990567818284, + 0.004470491781830788, + 0.0054458132945001125, + 0.003929086495190859, + 0.015676697716116905, + 0.029235074296593666, + 0.015954330563545227, + 0.02502477914094925, + 0.024390466511249542, + 0.016114074736833572, + 0.018954705446958542, + 0.029990598559379578, + 0.011567476205527782, + 0.013106195256114006, + 0.011652115732431412, + 0.010504364036023617, + 0.015537870116531849, + 0.02310042642056942, + 0.014561546966433525, + 0.016593515872955322, + 0.019056014716625214, + 0.017109785228967667, + 0.018682993948459625, + 0.02291836217045784, + 0.014473170973360538, + 0.027048690244555473, + 0.020137524232268333, + 0.01755853369832039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.0052520157769322395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, taken from a slightly low angle, captures a scene immersed in the surreal Makoto Shinkai aesthetic, the air itself brimming with emotion and light. The room is bathed in intense, golden late-afternoon sunlight, which filters through a large window draped with sheer white curtains, casting long, soft shadows and illuminating millions of dust particles dancing across the space in the visible twilight. On a light-colored wooden table lies a pristine stack of five vibrantly colored light novels, their glossy covers adorned with vibrant, saturated, anime-style character art. As if weightless, the entire volume rises smoothly and deliberately from the table's surface, a faint, ethereal aura shimmering beneath them. Once lifted into the air, the bookshelf gracefully parted, and each book began to gently rotate in the center of the room, a slow, enchanting ballet. The pages of the topmost book, with its vibrant pink and blue cover, began to flutter open, as if blown by a silent, magical breeze, revealing intricate illustrations from which emanated a brilliant, warm light, softly pulsing and casting a dynamic presence that shifted highlights throughout the room, creating a tranquil, breathtaking atmosphere.", + "seed": 12503, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_13/index_8480_seed_12503_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025063295662403107, + "vlm_logprob": -0.025063295662403107, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.025063295662403107, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003157750703394413 + ], + "opposite_logprob_tokens": [ + -8.003157615661621 + ], + "logprob_reward": -0.003157750703394413, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.003157750703394413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06157122924923897, + 0.03984696790575981, + 0.0798342153429985, + 0.030209917575120926, + 0.04354904219508171, + 0.052456650882959366, + 0.5977181792259216, + 0.03193653002381325, + 0.01751093380153179, + 0.04536627605557442 + ], + "credit_map": [ + 0.009071070700883865, + 0.012540710158646107, + 0.006921038031578064, + 0.006913403514772654, + 0.008302228525280952, + 0.006447287276387215, + 0.007203155197203159, + 0.004172331653535366, + 0.006145927123725414, + 0.007179794833064079, + 0.0035184870939701796, + 0.0030640990007668734, + 0.007192006800323725, + 0.006726735737174749, + 0.003450909163802862, + 0.002569010714069009, + 0.00781231140717864, + 0.008530329912900925, + 0.00453623291105032, + 0.004772337153553963, + 0.022821027785539627, + 0.02200772985816002, + 0.0052467333152890205, + 0.004107510205358267, + 0.003866910468786955, + 0.006268786732107401, + 0.0022281750570982695, + 0.0035425133537501097, + 0.005093990359455347, + 0.004380985628813505, + 0.0027706618420779705, + 0.0020578973926603794, + 0.00726306764408946, + 0.004653860814869404, + 0.0038755133282393217, + 0.005670030135661364, + 0.00574013777077198, + 0.007691266946494579, + 0.004406839143484831, + 0.004248325247317553, + 0.005846948362886906, + 0.003974458202719688, + 0.003909250721335411, + 0.004296234343200922, + 0.005753498058766127, + 0.013996308669447899, + 0.004508033394813538, + 0.010171915404498577, + 0.019509844481945038, + 0.01249248068779707, + 0.03173179179430008, + 0.020030725747346878, + 0.019428851082921028, + 0.2387571632862091, + 0.019730977714061737, + 0.23603637516498566, + 0.004179057665169239, + 0.0044695124961435795, + 0.0037416827399283648, + 0.00445341132581234, + 0.0033758182544261217, + 0.0036624774802476168, + 0.004533621482551098, + 0.003520944621413946, + 0.0020270736422389746, + 0.0015412640059366822, + 0.0017611476359888911, + 0.00325389695353806, + 0.0020142069552093744, + 0.0017772188875824213, + 0.003122767899185419, + 0.0020133578218519688, + 0.004859627690166235, + 0.004961075726896524, + 0.004786560777574778, + 0.006182701326906681, + 0.0064239720813930035, + 0.004268167540431023, + 0.005467970389872789, + 0.00841620285063982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.003157750703394413 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794038720428944 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.009794038720428944, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794038720428944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1138177439570427, + 0.06422816962003708, + 0.2236916720867157, + 0.032976843416690826, + 0.04460132494568825, + 0.03875362128019333, + 0.298714816570282, + 0.017793554812669754, + 0.11454761773347855, + 0.050874657928943634 + ], + "credit_map": [ + 0.017832007259130478, + 0.019839173182845116, + 0.013051177375018597, + 0.008177665993571281, + 0.019650040194392204, + 0.012983146123588085, + 0.011488686315715313, + 0.010795848444104195, + 0.01315913163125515, + 0.00784654263406992, + 0.007111277896910906, + 0.004419527482241392, + 0.013757904060184956, + 0.009806172922253609, + 0.004062422551214695, + 0.004065192770212889, + 0.020385121926665306, + 0.024316903203725815, + 0.011952335946261883, + 0.01609952189028263, + 0.06272685527801514, + 0.06134316325187683, + 0.016441959887742996, + 0.010425811633467674, + 0.0040100449696183205, + 0.005394575651735067, + 0.004387695342302322, + 0.004677860997617245, + 0.004264598712325096, + 0.004231493920087814, + 0.0035378760658204556, + 0.0024726998526602983, + 0.005277025979012251, + 0.0050672609359025955, + 0.0029551596380770206, + 0.0067780823446810246, + 0.006564467679709196, + 0.008386610075831413, + 0.005141329951584339, + 0.004431386943906546, + 0.008286843076348305, + 0.0036940122954547405, + 0.003395922714844346, + 0.0027203168720006943, + 0.0036485313903540373, + 0.008245292119681835, + 0.002960569690912962, + 0.005802130792289972, + 0.012257932685315609, + 0.011265589855611324, + 0.01957392506301403, + 0.02093665488064289, + 0.014470300637185574, + 0.0990627184510231, + 0.022344689816236496, + 0.09880301356315613, + 0.002966703847050667, + 0.0016940346686169505, + 0.0020396222826093435, + 0.0025164373219013214, + 0.0019771037623286247, + 0.002017229562625289, + 0.0018435458187013865, + 0.0027388762682676315, + 0.009295757859945297, + 0.026370888575911522, + 0.014105398207902908, + 0.007412504870444536, + 0.013344241306185722, + 0.015606281347572803, + 0.007328708190470934, + 0.02108384296298027, + 0.005184582434594631, + 0.005734823178499937, + 0.006245611235499382, + 0.004694701172411442, + 0.0054622613824903965, + 0.006500863004475832, + 0.0069014038890600204, + 0.010150410234928131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.009794038720428944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554306536912918 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.012554306536912918, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554306536912918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06790031492710114, + 0.026210280135273933, + 0.46357443928718567, + 0.02717510610818863, + 0.08068681508302689, + 0.08037009835243225, + 0.1211744174361229, + 0.020309094339609146, + 0.07067733258008957, + 0.041922155767679214 + ], + "credit_map": [ + 0.008181081153452396, + 0.007301973178982735, + 0.007295200135558844, + 0.011589187197387218, + 0.007166984491050243, + 0.011243791319429874, + 0.008877887390553951, + 0.006244199350476265, + 0.00332453940063715, + 0.003199380123987794, + 0.0027086883783340454, + 0.003877694020047784, + 0.003711973549798131, + 0.0026419786736369133, + 0.004524646792560816, + 0.00222137663513422, + 0.04089614748954773, + 0.052010636776685715, + 0.02400345355272293, + 0.03493969142436981, + 0.1286696195602417, + 0.12459881603717804, + 0.03563744202256203, + 0.022818587720394135, + 0.003999272361397743, + 0.005087635479867458, + 0.0018563277553766966, + 0.00345232174731791, + 0.004193588625639677, + 0.0032891673035919666, + 0.0032326269429177046, + 0.0020641633309423923, + 0.016646282747387886, + 0.007869192399084568, + 0.007928241975605488, + 0.009342278353869915, + 0.011245672591030598, + 0.009270241484045982, + 0.009538833051919937, + 0.008846061304211617, + 0.009149609133601189, + 0.006930318661034107, + 0.009212331846356392, + 0.00956158060580492, + 0.010277136228978634, + 0.01669052243232727, + 0.010153507813811302, + 0.008395091630518436, + 0.00466707069426775, + 0.003406073898077011, + 0.007132600527256727, + 0.005183661822229624, + 0.0048493891954422, + 0.04568185284733772, + 0.005587775260210037, + 0.04466598108410835, + 0.003621551673859358, + 0.0018882086733356118, + 0.003231905633583665, + 0.002639180049300194, + 0.0018526401836425066, + 0.0021767369471490383, + 0.002712355460971594, + 0.0021865139715373516, + 0.004906188230961561, + 0.017345258966088295, + 0.006221185903996229, + 0.004939935635775328, + 0.0074820793233811855, + 0.005334608722478151, + 0.005491271149367094, + 0.01895679533481598, + 0.004547025542706251, + 0.003794027492403984, + 0.004227265249937773, + 0.005564188119024038, + 0.006695244926959276, + 0.0042824577540159225, + 0.007514275144785643, + 0.0052976650185883045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.012554306536912918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001148164272308 + ], + "opposite_logprob_tokens": [ + -2.6950113773345947 + ], + "logprob_reward": -0.07001148164272308, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001148164272308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.089670330286026, + 0.040691010653972626, + 0.06365679949522018, + 0.04569956660270691, + 0.10539120435714722, + 0.057770635932683945, + 0.29778361320495605, + 0.042160823941230774, + 0.2203277051448822, + 0.036848317831754684 + ], + "credit_map": [ + 0.014317325316369534, + 0.011424886994063854, + 0.010107632726430893, + 0.010397456586360931, + 0.01305375061929226, + 0.011784449219703674, + 0.009149610064923763, + 0.009435220621526241, + 0.005040164105594158, + 0.004557616077363491, + 0.0033223123755306005, + 0.004267133306711912, + 0.006560319568961859, + 0.00728004751726985, + 0.00590138603001833, + 0.0037620337679982185, + 0.005997618660330772, + 0.007572683040052652, + 0.004598682280629873, + 0.005199901759624481, + 0.014681047759950161, + 0.01513223722577095, + 0.0061719841323792934, + 0.004302642308175564, + 0.007352855056524277, + 0.006812706124037504, + 0.005528565496206284, + 0.005199282895773649, + 0.005234184209257364, + 0.0070280227810144424, + 0.004012869205325842, + 0.004531081300228834, + 0.02176009677350521, + 0.01133800484240055, + 0.011064035817980766, + 0.010968675836920738, + 0.012677207589149475, + 0.018558107316493988, + 0.01060792338103056, + 0.00841715931892395, + 0.009789442643523216, + 0.004466358572244644, + 0.004809722304344177, + 0.005764725152403116, + 0.0059039536863565445, + 0.012086273171007633, + 0.005614155437797308, + 0.00933600403368473, + 0.010618721134960651, + 0.00710641173645854, + 0.01627269573509693, + 0.013359601609408855, + 0.01189514435827732, + 0.11363862454891205, + 0.01399792917072773, + 0.11089449375867844, + 0.006707291584461927, + 0.0030885019805282354, + 0.007707678247243166, + 0.005723117850720882, + 0.0029095395002514124, + 0.0045762863010168076, + 0.0057528335601091385, + 0.005695573519915342, + 0.011659301817417145, + 0.061147067695856094, + 0.016836851835250854, + 0.016456013545393944, + 0.02536091022193432, + 0.014585074037313461, + 0.019976157695055008, + 0.05430632829666138, + 0.0036646516527980566, + 0.003262205282226205, + 0.004049991257488728, + 0.005865745712071657, + 0.005297234281897545, + 0.003952521365135908, + 0.0059134080074727535, + 0.0048425584100186825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.07001148164272308 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798900708556175 + ], + "opposite_logprob_tokens": [ + -3.529798984527588 + ], + "logprob_reward": -0.029798900708556175, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798900708556175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08752848953008652, + 0.036701757460832596, + 0.25356271862983704, + 0.07707548886537552, + 0.050662849098443985, + 0.08979451656341553, + 0.056016705930233, + 0.06313758343458176, + 0.16492491960525513, + 0.12059491872787476 + ], + "credit_map": [ + 0.014643522910773754, + 0.01413731649518013, + 0.009504282847046852, + 0.00748864933848381, + 0.013446909375488758, + 0.010391595773398876, + 0.008167230524122715, + 0.00974898412823677, + 0.004064256325364113, + 0.004430097993463278, + 0.0029074805788695812, + 0.002617195248603821, + 0.009478743188083172, + 0.007776467129588127, + 0.0024407864548265934, + 0.002986731706187129, + 0.022566741332411766, + 0.02735290862619877, + 0.015504852868616581, + 0.017945803701877594, + 0.07197405397891998, + 0.06909007579088211, + 0.018104467540979385, + 0.011023844592273235, + 0.008943342603743076, + 0.008944827131927013, + 0.0057509178295731544, + 0.005557645577937365, + 0.020668216049671173, + 0.017265262082219124, + 0.0052876221016049385, + 0.0046576643362641335, + 0.009569764137268066, + 0.00696386955678463, + 0.0058453269302845, + 0.0036502955481410027, + 0.007402088493108749, + 0.008350265212357044, + 0.004458032548427582, + 0.004423215519636869, + 0.009407849051058292, + 0.008244448341429234, + 0.008369382470846176, + 0.009525866247713566, + 0.009092006832361221, + 0.014859763905405998, + 0.013448052108287811, + 0.01684715412557125, + 0.003603392979130149, + 0.004580586217343807, + 0.0060070171020925045, + 0.004263808485120535, + 0.003910341300070286, + 0.014876012690365314, + 0.005034167319536209, + 0.013741384260356426, + 0.005977412220090628, + 0.0062217749655246735, + 0.008221824653446674, + 0.0104450648650527, + 0.005666307173669338, + 0.00854167528450489, + 0.010587573051452637, + 0.007475954946130514, + 0.009190995246171951, + 0.04512565955519676, + 0.012883630581200123, + 0.013512767851352692, + 0.018688274547457695, + 0.011645393446087837, + 0.014691685326397419, + 0.039186518639326096, + 0.01073887012898922, + 0.011176708154380322, + 0.012022112496197224, + 0.012799572199583054, + 0.017208842560648918, + 0.013365528546273708, + 0.017747072502970695, + 0.02553621493279934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.029798900708556175 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, taken from a slightly low angle, captures a scene immersed in the surreal Makoto Shinkai aesthetic, the air itself brimming with emotion and light. The room is bathed in intense, golden late-afternoon sunlight, which filters through a large window draped with sheer white curtains, casting long, soft shadows and illuminating millions of dust particles dancing across the space in the visible twilight. On a light-colored wooden table lies a pristine stack of five vibrantly colored light novels, their glossy covers adorned with vibrant, saturated, anime-style character art. As if weightless, the entire volume rises smoothly and deliberately from the table's surface, a faint, ethereal aura shimmering beneath them. Once lifted into the air, the bookshelf gracefully parted, and each book began to gently rotate in the center of the room, a slow, enchanting ballet. The pages of the topmost book, with its vibrant pink and blue cover, began to flutter open, as if blown by a silent, magical breeze, revealing intricate illustrations from which emanated a brilliant, warm light, softly pulsing and casting a dynamic presence that shifted highlights throughout the room, creating a tranquil, breathtaking atmosphere.", + "seed": 12504, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_13/index_8480_seed_12504_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026791885076090694, + "vlm_logprob": -0.026791885076090694, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.026791885076090694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06571916490793228 + ], + "opposite_logprob_tokens": [ + -2.8157191276550293 + ], + "logprob_reward": -0.06571916490793228, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06571916490793228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12340860068798065, + 0.075706347823143, + 0.25711071491241455, + 0.058194469660520554, + 0.11773748695850372, + 0.10179764032363892, + 0.14880506694316864, + 0.040206655859947205, + 0.02493748627603054, + 0.052095554769039154 + ], + "credit_map": [ + 0.022090185433626175, + 0.013504073955118656, + 0.013260426931083202, + 0.015384005382657051, + 0.014071271754801273, + 0.009328490123152733, + 0.020574502646923065, + 0.015195636078715324, + 0.01647222228348255, + 0.005976444575935602, + 0.011166519485414028, + 0.005876152776181698, + 0.005297754891216755, + 0.01165792252868414, + 0.007797021884471178, + 0.01146230474114418, + 0.032776568084955215, + 0.03295873478055, + 0.036612093448638916, + 0.034688763320446014, + 0.019942233338952065, + 0.03768002986907959, + 0.0354241281747818, + 0.0270281583070755, + 0.007965482771396637, + 0.003898884169757366, + 0.011182783171534538, + 0.006887311115860939, + 0.004252214916050434, + 0.00793003011494875, + 0.008581243455410004, + 0.007496512494981289, + 0.01860407181084156, + 0.010512316599488258, + 0.014531021937727928, + 0.010397951118648052, + 0.015066305175423622, + 0.015333939343690872, + 0.009749107994139194, + 0.02354276180267334, + 0.018476754426956177, + 0.012231244705617428, + 0.019369330257177353, + 0.020393991842865944, + 0.011140234768390656, + 0.006768820341676474, + 0.004908302798867226, + 0.008508956991136074, + 0.025496242567896843, + 0.04582897573709488, + 0.02622343972325325, + 0.011314218863844872, + 0.016117453575134277, + 0.005435440223664045, + 0.00782101135700941, + 0.010568279772996902, + 0.004269011784344912, + 0.0054667494259774685, + 0.004880973137915134, + 0.005404922645539045, + 0.008270544931292534, + 0.003521509235724807, + 0.005032382905483246, + 0.0033605608623474836, + 0.0025547812692821026, + 0.003212256822735071, + 0.0026063055265694857, + 0.004187033977359533, + 0.004805344622582197, + 0.0027303670067340136, + 0.0022039737086743116, + 0.002637421013787389, + 0.0046460265293717384, + 0.007989075034856796, + 0.003762920619919896, + 0.006365359295159578, + 0.009925706312060356, + 0.00335843488574028, + 0.006571277044713497, + 0.009476752951741219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.06571916490793228 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010624246206134558 + ], + "opposite_logprob_tokens": [ + -6.876062393188477 + ], + "logprob_reward": -0.0010624246206134558, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010624246206134558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09778311848640442, + 0.1647965908050537, + 0.08295967429876328, + 0.028509918600320816, + 0.0489162802696228, + 0.04763687402009964, + 0.2698250710964203, + 0.13021555542945862, + 0.055876173079013824, + 0.0734807550907135 + ], + "credit_map": [ + 0.01422057393938303, + 0.011663916520774364, + 0.01239177118986845, + 0.015184380114078522, + 0.009130909107625484, + 0.007284932304173708, + 0.016118284314870834, + 0.011788349598646164, + 0.03442823886871338, + 0.018592627719044685, + 0.034003738313913345, + 0.013545917347073555, + 0.007407897152006626, + 0.015558730810880661, + 0.015556594356894493, + 0.025702835991978645, + 0.01085437461733818, + 0.01590709760785103, + 0.008670310489833355, + 0.013627278618514538, + 0.007196144666522741, + 0.008544974960386753, + 0.009681093506515026, + 0.00847840029746294, + 0.0043352278880774975, + 0.001801579026505351, + 0.004268779885023832, + 0.002915230579674244, + 0.0030514178797602654, + 0.0029764024075120687, + 0.005261664278805256, + 0.0038996157236397266, + 0.011889404617249966, + 0.004964597523212433, + 0.0052454727701842785, + 0.0047658393159508705, + 0.0030615348368883133, + 0.004307182040065527, + 0.004628659226000309, + 0.010053588077425957, + 0.006288307253271341, + 0.006081883795559406, + 0.004959942307323217, + 0.008510119281709194, + 0.005500080529600382, + 0.004291456192731857, + 0.007496897131204605, + 0.00450819032266736, + 0.045605454593896866, + 0.08352221548557281, + 0.05438920482993126, + 0.01886657439172268, + 0.023215217515826225, + 0.010741300880908966, + 0.012218831107020378, + 0.021266264840960503, + 0.016161972656846046, + 0.014181680046021938, + 0.02458147145807743, + 0.020740076899528503, + 0.024056384339928627, + 0.009122274816036224, + 0.010296381078660488, + 0.011075315065681934, + 0.006610617507249117, + 0.008017288520932198, + 0.007452783174812794, + 0.006282793823629618, + 0.012590952217578888, + 0.004037581849843264, + 0.006338033825159073, + 0.004546123091131449, + 0.008284940384328365, + 0.010366545058786869, + 0.0061814384534955025, + 0.012046665884554386, + 0.013823114335536957, + 0.004476039670407772, + 0.0074905697256326675, + 0.01081144530326128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.0010624246206134558 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020599940791726112 + ], + "opposite_logprob_tokens": [ + -3.895599842071533 + ], + "logprob_reward": -0.020599940791726112, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020599940791726112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05289650335907936, + 0.17680156230926514, + 0.09028248488903046, + 0.08132453262805939, + 0.07351270318031311, + 0.07850658893585205, + 0.17778530716896057, + 0.16937892138957977, + 0.07112191617488861, + 0.028389442712068558 + ], + "credit_map": [ + 0.006324563641101122, + 0.00582691328600049, + 0.004560871981084347, + 0.0071575031615793705, + 0.005722541827708483, + 0.0032638374250382185, + 0.013873458839952946, + 0.006166817620396614, + 0.017006995156407356, + 0.011389730498194695, + 0.01643471047282219, + 0.016246771439909935, + 0.020520083606243134, + 0.016395090147852898, + 0.05730723217129707, + 0.02150096744298935, + 0.012891259044408798, + 0.008161825127899647, + 0.013015957549214363, + 0.010929897427558899, + 0.006771738175302744, + 0.01418008841574192, + 0.011594390496611595, + 0.012737328186631203, + 0.008428295142948627, + 0.0069509283639490604, + 0.007600302807986736, + 0.010944399051368237, + 0.010323015972971916, + 0.0075905295088887215, + 0.01994278095662594, + 0.00954427756369114, + 0.014444381929934025, + 0.007262853439897299, + 0.007057459093630314, + 0.00787336565554142, + 0.009417814202606678, + 0.006430394947528839, + 0.009242307394742966, + 0.01178413163870573, + 0.0106047373265028, + 0.011701976880431175, + 0.01285820733755827, + 0.008032888174057007, + 0.01208462007343769, + 0.00814513023942709, + 0.00850172620266676, + 0.006577303167432547, + 0.019306182861328125, + 0.03160823509097099, + 0.029841026291251183, + 0.017349345609545708, + 0.02667737565934658, + 0.022023893892765045, + 0.013719987124204636, + 0.017259279265999794, + 0.020983781665563583, + 0.02456473559141159, + 0.010275756008923054, + 0.020286530256271362, + 0.056108251214027405, + 0.013326235115528107, + 0.015183748677372932, + 0.008649894967675209, + 0.006556736771017313, + 0.010836266912519932, + 0.0073035345412790775, + 0.007464130874723196, + 0.018444104120135307, + 0.0061277667991817, + 0.0073919096030294895, + 0.006997466087341309, + 0.0035880273208022118, + 0.004369521047919989, + 0.003974025137722492, + 0.0030077078845351934, + 0.006130114663392305, + 0.0019402150064706802, + 0.0029904458206146955, + 0.0023893883917480707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.020599940791726112 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287607356905937 + ], + "opposite_logprob_tokens": [ + -3.773287534713745 + ], + "logprob_reward": -0.023287607356905937, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287607356905937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.085745669901371, + 0.07337537407875061, + 0.050098713487386703, + 0.03357486426830292, + 0.03692412003874779, + 0.06306729465723038, + 0.4574289321899414, + 0.12811745703220367, + 0.035589393228292465, + 0.036078207194805145 + ], + "credit_map": [ + 0.013744289055466652, + 0.010102632455527782, + 0.010482128709554672, + 0.011282491497695446, + 0.00705022644251585, + 0.007233044132590294, + 0.013563789427280426, + 0.01228706631809473, + 0.008488290011882782, + 0.006617797538638115, + 0.010662646032869816, + 0.007037663366645575, + 0.006319783627986908, + 0.01405609492212534, + 0.009244087152183056, + 0.010949010960757732, + 0.006290723569691181, + 0.008474303409457207, + 0.0068877628073096275, + 0.005263684783130884, + 0.003760751336812973, + 0.006657432299107313, + 0.007208818104118109, + 0.0055552395060658455, + 0.004023557063192129, + 0.002625170163810253, + 0.003267961088567972, + 0.004398464225232601, + 0.004266498610377312, + 0.004766764584928751, + 0.0066109513863921165, + 0.003615495515987277, + 0.004431874491274357, + 0.0036791828460991383, + 0.0038938771467655897, + 0.005547796841710806, + 0.006600250955671072, + 0.0033537454437464476, + 0.0055029806680977345, + 0.003914411645382643, + 0.007784527260810137, + 0.007403508760035038, + 0.012245078571140766, + 0.007676661480218172, + 0.008546603843569756, + 0.0062450082041323185, + 0.007067426107823849, + 0.00609848415479064, + 0.08298686891794205, + 0.1508135050535202, + 0.08686605095863342, + 0.0317874550819397, + 0.039377786219120026, + 0.015941891819238663, + 0.014204814098775387, + 0.0354505330324173, + 0.01268882304430008, + 0.014616170898079872, + 0.012266364879906178, + 0.024414533749222755, + 0.021660391241312027, + 0.015981975942850113, + 0.01617022603750229, + 0.010318972170352936, + 0.0034111705608665943, + 0.004940458573400974, + 0.0036468945909291506, + 0.0049734474159777164, + 0.008143939077854156, + 0.0033844932913780212, + 0.004242453724145889, + 0.0028465348295867443, + 0.0031391356606036425, + 0.0055258264765143394, + 0.0034441102761775255, + 0.003808927023783326, + 0.007799070328474045, + 0.003302669385448098, + 0.005692279897630215, + 0.0033661872148513794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.023287607356905937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a room with a large window draped with sheer white curtains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the room illuminated by intense golden late-afternoon sunlight filtering through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stack of five vibrantly colored light novels with anime-style character art visible on a light-colored wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the books rise smoothly from the table and then rotate gently in the center of the room?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the topmost book with a pink and blue cover flutter open to reveal illustrations emitting a brilliant, warm, pulsing light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329028770327568 + ], + "opposite_logprob_tokens": [ + -3.7732903957366943 + ], + "logprob_reward": -0.02329028770327568, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329028770327568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05379411578178406, + 0.07111823558807373, + 0.028822170570492744, + 0.03825875744223595, + 0.09310172498226166, + 0.09298421442508698, + 0.10812738537788391, + 0.17771026492118835, + 0.19972462952136993, + 0.13635848462581635 + ], + "credit_map": [ + 0.006496117450296879, + 0.006166165228933096, + 0.004520058631896973, + 0.007162724621593952, + 0.006247465964406729, + 0.0035566247534006834, + 0.013504155911505222, + 0.006140811834484339, + 0.010676159523427486, + 0.007766137830913067, + 0.014345694333314896, + 0.005668077617883682, + 0.004153009038418531, + 0.0059023830108344555, + 0.011486472561955452, + 0.011120310053229332, + 0.004560226108878851, + 0.0019260369008406997, + 0.004148881882429123, + 0.0032814948353916407, + 0.0022187214344739914, + 0.003359995549544692, + 0.005482527893036604, + 0.0038442867808043957, + 0.004214148502796888, + 0.003913812339305878, + 0.0056400420144200325, + 0.00440822821110487, + 0.00519555015489459, + 0.0034704627469182014, + 0.006881118286401033, + 0.004535399843007326, + 0.015084635466337204, + 0.007350336294621229, + 0.010746114887297153, + 0.007668281905353069, + 0.013550476171076298, + 0.010999217629432678, + 0.01027589663863182, + 0.01742677018046379, + 0.011452557519078255, + 0.013943752273917198, + 0.013423155061900616, + 0.014510772190988064, + 0.007452613674104214, + 0.00797571986913681, + 0.01347188651561737, + 0.010753756389021873, + 0.01082263607531786, + 0.01818017102777958, + 0.019278645515441895, + 0.012195616960525513, + 0.01684620790183544, + 0.008228220976889133, + 0.013088498264551163, + 0.009487404488027096, + 0.017463305965065956, + 0.023885833099484444, + 0.016835477203130722, + 0.030773987993597984, + 0.038335170596838, + 0.018232271075248718, + 0.02065541036427021, + 0.011528818868100643, + 0.01753328926861286, + 0.028849104419350624, + 0.0243399478495121, + 0.02850313112139702, + 0.04225914180278778, + 0.018325675278902054, + 0.023732269182801247, + 0.016182074323296547, + 0.009251619689166546, + 0.023548142984509468, + 0.010237721726298332, + 0.01739504374563694, + 0.038729626685380936, + 0.007155874278396368, + 0.016348544508218765, + 0.013691914267838001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.02329028770327568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_14/index_tensor([1938, 1938])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_14/index_tensor([1938, 1938])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c0de59ca16599fc1737deb596c7e72b505724cc3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_14/index_tensor([1938, 1938])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, high-angle aerial shot gracefully follows and slowly circles a majestic eagle, its massive, dark brown wings fully extended, effortlessly riding the high-altitude thermals. In the clear, high-altitude sunlight, the edges of each feather are rimmed with a golden hue, with the signature golden-brown feathers on the nape of its neck being particularly striking. The eagle maintains near-perfect stability, with only the wingtip feathers making minute, precise adjustments. Its head methodically turns from side to side, its amber eyes, sharp as glass beads, scanning the majestic mountains thousands of feet below with cool, focused attention. Below, an awe-inspiring panorama of rugged mountains stretches endlessly below, their peaks capped with snow, while deep green coniferous forests carpet the slopes like a thick carpet, and the exposed gray rock faces reveal rich textures in the light. A winding river, like a silver ribbon, weaves through a deep canyon, its waters reflecting dazzling light. Set against the backdrop of a vast blue sky, the entire image is imbued with a sense of tranquility and powerful energy, conveying the pristine, undisturbed beauty of nature and the solitary majesty of a top predator.", + "seed": 7845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_14/index_1938_seed_7845_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.032722073746845125, + "vlm_logprob": -0.032722073746845125, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.032722073746845125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0166845191270113 + ], + "opposite_logprob_tokens": [ + -4.266684532165527 + ], + "logprob_reward": -0.0166845191270113, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0166845191270113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059643130749464035, + 0.03594250977039337, + 0.05201980099081993, + 0.040221285074949265, + 0.05966983735561371, + 0.20987990498542786, + 0.2721176743507385, + 0.06730782240629196, + 0.026908203959465027, + 0.17628982663154602 + ], + "credit_map": [ + 0.007428656332194805, + 0.004257214721292257, + 0.005937560927122831, + 0.010754789225757122, + 0.005989057943224907, + 0.0050523532554507256, + 0.010526555590331554, + 0.009696947410702705, + 0.00411281967535615, + 0.0018974777776747942, + 0.006498384289443493, + 0.010204211808741093, + 0.0023646089248359203, + 0.0023628852795809507, + 0.004113652743399143, + 0.0043884688057005405, + 0.005339133553206921, + 0.0034605362452566624, + 0.011137028224766254, + 0.01082634087651968, + 0.0035662094596773386, + 0.0029107059817761183, + 0.00728602847084403, + 0.007493817247450352, + 0.00409966753795743, + 0.0023420038633048534, + 0.00868330430239439, + 0.00732968607917428, + 0.0026672547683119774, + 0.0027577532455325127, + 0.007346530444920063, + 0.004995083902031183, + 0.005718914791941643, + 0.003122675698250532, + 0.011861782521009445, + 0.008992102928459644, + 0.005400959402322769, + 0.006960704922676086, + 0.00854580756276846, + 0.009066890925168991, + 0.014402729459106922, + 0.024394208565354347, + 0.03424818068742752, + 0.037656914442777634, + 0.013673949055373669, + 0.017057137563824654, + 0.03517630696296692, + 0.03327048569917679, + 0.019653310999274254, + 0.02300955355167389, + 0.08521386981010437, + 0.06243149936199188, + 0.01186302863061428, + 0.012686730362474918, + 0.026253828778862953, + 0.031005874276161194, + 0.007133823353797197, + 0.005114974919706583, + 0.011738840490579605, + 0.014738243073225021, + 0.004663677886128426, + 0.006472358945757151, + 0.010064907371997833, + 0.007380994036793709, + 0.0027043188456445932, + 0.0016275257803499699, + 0.005940363276749849, + 0.00453184312209487, + 0.0023902098182588816, + 0.0030303115490823984, + 0.0025495709851384163, + 0.004134057555347681, + 0.01812003180384636, + 0.014243140816688538, + 0.023015804588794708, + 0.02133421041071415, + 0.015302407555282116, + 0.016540277749300003, + 0.034080345183610916, + 0.033653624355793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.0166845191270113 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020599357783794403 + ], + "opposite_logprob_tokens": [ + -3.895599365234375 + ], + "logprob_reward": -0.020599357783794403, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020599357783794403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12095662951469421, + 0.03306686505675316, + 0.07010014355182648, + 0.10079608112573624, + 0.03647040203213692, + 0.12027749419212341, + 0.062237322330474854, + 0.26763421297073364, + 0.0592983178794384, + 0.1291625201702118 + ], + "credit_map": [ + 0.010733471252024174, + 0.007820436730980873, + 0.011756722815334797, + 0.015768295153975487, + 0.01584050990641117, + 0.009238787926733494, + 0.03159968927502632, + 0.018198708072304726, + 0.004997855052351952, + 0.002277924446389079, + 0.004175293724983931, + 0.008494959212839603, + 0.0017924547428265214, + 0.002577331615611911, + 0.004552698228508234, + 0.004198348615318537, + 0.007082900032401085, + 0.005335325840860605, + 0.014381708577275276, + 0.014569424092769623, + 0.0027915961109101772, + 0.005605428013950586, + 0.011349979788064957, + 0.008983778767287731, + 0.009926906786859035, + 0.004279209766536951, + 0.03999891132116318, + 0.02243483066558838, + 0.0034233606420457363, + 0.0035841588396579027, + 0.009435656480491161, + 0.007713040336966515, + 0.0037248549051582813, + 0.0021070430520921946, + 0.006010768935084343, + 0.008227096870541573, + 0.0030813675839453936, + 0.0033072878140956163, + 0.005318277049809694, + 0.004693706054240465, + 0.007604725193232298, + 0.006670158822089434, + 0.042695894837379456, + 0.026287559419870377, + 0.0041732266545295715, + 0.005707805044949055, + 0.01823986880481243, + 0.008898250758647919, + 0.004282393492758274, + 0.004738191608339548, + 0.017622068524360657, + 0.012882432900369167, + 0.0034124809317290783, + 0.00555361807346344, + 0.006427816115319729, + 0.0073183211497962475, + 0.02427968569099903, + 0.012863444164395332, + 0.05697111040353775, + 0.04513029754161835, + 0.018595170229673386, + 0.04026082903146744, + 0.031986549496650696, + 0.03754713758826256, + 0.007650529500097036, + 0.003612757893279195, + 0.015627378597855568, + 0.010160746984183788, + 0.004128816071897745, + 0.004344229586422443, + 0.006201821845024824, + 0.007572034373879433, + 0.01882982812821865, + 0.009438049979507923, + 0.017803605645895004, + 0.021320778876543045, + 0.009016952477395535, + 0.01028679870069027, + 0.024571184068918228, + 0.017895331606268883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.020599357783794403 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129349023103714 + ], + "opposite_logprob_tokens": [ + -2.2379348278045654 + ], + "logprob_reward": -0.1129349023103714, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129349023103714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08301342278718948, + 0.028690647333860397, + 0.11285393685102463, + 0.052700240164995193, + 0.06812513619661331, + 0.10701613873243332, + 0.27716290950775146, + 0.05266517773270607, + 0.060527265071868896, + 0.15724515914916992 + ], + "credit_map": [ + 0.007607096340507269, + 0.006056440528482199, + 0.007360576186329126, + 0.009891237132251263, + 0.011029019020497799, + 0.008009597659111023, + 0.020961763337254524, + 0.012097680009901524, + 0.004126302897930145, + 0.002416482660919428, + 0.0037532614078372717, + 0.006715431809425354, + 0.001712016761302948, + 0.0019759521819651127, + 0.004002291709184647, + 0.003988904412835836, + 0.01073271594941616, + 0.005205465015023947, + 0.025673096999526024, + 0.0211564302444458, + 0.0075295791029930115, + 0.009611674584448338, + 0.01729496195912361, + 0.015650002285838127, + 0.0034766800235956907, + 0.0018982746405526996, + 0.020245075225830078, + 0.011972215957939625, + 0.0019865259528160095, + 0.002731712767854333, + 0.004993933252990246, + 0.0053958166390657425, + 0.005138423293828964, + 0.003296040464192629, + 0.012256314046680927, + 0.020390786230564117, + 0.004665185231715441, + 0.006563397590070963, + 0.008935797959566116, + 0.0068791876547038555, + 0.0057647074572741985, + 0.0068184733390808105, + 0.02962449938058853, + 0.023064367473125458, + 0.005419878754764795, + 0.006239637732505798, + 0.016198568046092987, + 0.013886000961065292, + 0.020914649590849876, + 0.02445666678249836, + 0.08393645286560059, + 0.056753139942884445, + 0.012521877884864807, + 0.014743120409548283, + 0.028098132461309433, + 0.03573886305093765, + 0.005480730906128883, + 0.0032576839439570904, + 0.008946175687015057, + 0.010899204760789871, + 0.003618015209212899, + 0.0054096742533147335, + 0.009157529100775719, + 0.005896164104342461, + 0.004458970855921507, + 0.003758287988603115, + 0.009181595407426357, + 0.011852634139358997, + 0.005222110543400049, + 0.007379042915999889, + 0.007770666852593422, + 0.010903948917984962, + 0.02460004761815071, + 0.012974703684449196, + 0.023089874535799026, + 0.019910214468836784, + 0.01328397449105978, + 0.01478191465139389, + 0.014608007855713367, + 0.033996399492025375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.1129349023103714 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778928942978382 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778928942978382, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778928942978382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23207052052021027, + 0.1496618688106537, + 0.06404528766870499, + 0.10282071679830551, + 0.026241669431328773, + 0.0603187121450901, + 0.2045401781797409, + 0.054312240332365036, + 0.04010886326432228, + 0.06587997078895569 + ], + "credit_map": [ + 0.020909463986754417, + 0.015228834003210068, + 0.022924991324543953, + 0.03171509504318237, + 0.028712695464491844, + 0.02589298225939274, + 0.0540865920484066, + 0.032599885016679764, + 0.014419786632061005, + 0.0066436417400836945, + 0.033224917948246, + 0.04816634580492973, + 0.00755701819434762, + 0.0071256570518016815, + 0.014141875319182873, + 0.01838262565433979, + 0.005569531582295895, + 0.0040137446485459805, + 0.011884582228958607, + 0.011100425384938717, + 0.003249820088967681, + 0.004678643308579922, + 0.012499871663749218, + 0.011048675514757633, + 0.006675957702100277, + 0.0037936856970191, + 0.04399636387825012, + 0.025912906974554062, + 0.002690780209377408, + 0.003129328368231654, + 0.009116573259234428, + 0.0075051323510706425, + 0.00226770481094718, + 0.0015101705212146044, + 0.00493330042809248, + 0.006351635791361332, + 0.0018261197255924344, + 0.0015974633861333132, + 0.0032906834967434406, + 0.0044645932503044605, + 0.004357052966952324, + 0.004042196087539196, + 0.01736639440059662, + 0.010402155108749866, + 0.004001985304057598, + 0.004296513739973307, + 0.009532283991575241, + 0.0063201310113072395, + 0.016279367730021477, + 0.01495374459773302, + 0.06256169825792313, + 0.03842585161328316, + 0.012216144241392612, + 0.01561204344034195, + 0.021611148491501808, + 0.022880177944898605, + 0.006245729513466358, + 0.0038159526884555817, + 0.009497379884123802, + 0.010260378010571003, + 0.0028737331740558147, + 0.0040620132349431515, + 0.010605890303850174, + 0.006951167248189449, + 0.003755016252398491, + 0.0022218734957277775, + 0.006639743223786354, + 0.006298088002949953, + 0.004062999971210957, + 0.004474314860999584, + 0.005014830268919468, + 0.007642001379281282, + 0.005639408249408007, + 0.006440021563321352, + 0.010795163922011852, + 0.015106264501810074, + 0.004603452980518341, + 0.005354658234864473, + 0.008491125889122486, + 0.009449874050915241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.009778928942978382 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036126605700701475 + ], + "opposite_logprob_tokens": [ + -5.628612518310547 + ], + "logprob_reward": -0.0036126605700701475, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036126605700701475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07857110351324081, + 0.05110827833414078, + 0.09523222595453262, + 0.03961508348584175, + 0.08569762110710144, + 0.11535900086164474, + 0.17633850872516632, + 0.10955540090799332, + 0.055263109505176544, + 0.19325971603393555 + ], + "credit_map": [ + 0.009619480930268764, + 0.004744073376059532, + 0.007725048810243607, + 0.0127249201759696, + 0.008773144334554672, + 0.0068603320978581905, + 0.018363654613494873, + 0.009760437533259392, + 0.006346424575895071, + 0.0034611925948411226, + 0.006059772334992886, + 0.011399375274777412, + 0.0049521829932928085, + 0.005100781563669443, + 0.006784266326576471, + 0.007004274521023035, + 0.009114698506891727, + 0.006103401072323322, + 0.024672238156199455, + 0.0170445553958416, + 0.0050940969958901405, + 0.005397085566073656, + 0.01611221209168434, + 0.011693933978676796, + 0.003757015336304903, + 0.0023482805117964745, + 0.0060129654593765736, + 0.010132781229913235, + 0.0025942986831068993, + 0.0018580054165795445, + 0.008684022352099419, + 0.004227710422128439, + 0.007774244993925095, + 0.0035021475050598383, + 0.015449333935976028, + 0.021634042263031006, + 0.004639806691557169, + 0.0065182847902178764, + 0.011309840716421604, + 0.01486990787088871, + 0.00877830758690834, + 0.00809735618531704, + 0.030073489993810654, + 0.023094328120350838, + 0.005878958385437727, + 0.007295810151845217, + 0.019061999395489693, + 0.013078739866614342, + 0.015903232619166374, + 0.01097539160400629, + 0.05925898253917694, + 0.03894717991352081, + 0.007042580749839544, + 0.011524676345288754, + 0.015390272252261639, + 0.017296183854341507, + 0.012615512125194073, + 0.006471265573054552, + 0.025029536336660385, + 0.01977783441543579, + 0.006313868798315525, + 0.013117981143295765, + 0.013408366590738297, + 0.012821020558476448, + 0.0050668721087276936, + 0.0031382206361740828, + 0.0099503044039011, + 0.012316263280808926, + 0.00377374398522079, + 0.0043304250575602055, + 0.00905237253755331, + 0.007634901907294989, + 0.021785402670502663, + 0.014304733835160732, + 0.02488245628774166, + 0.026766633614897728, + 0.016511620953679085, + 0.01926986128091812, + 0.023284858092665672, + 0.046454139053821564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.0036126605700701475 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle aerial shot gracefully follows and slowly circles a majestic eagle, its massive, dark brown wings fully extended, effortlessly riding the high-altitude thermals. In the clear, high-altitude sunlight, the edges of each feather are rimmed with a golden hue, with the signature golden-brown feathers on the nape of its neck being particularly striking. The eagle maintains near-perfect stability, with only the wingtip feathers making minute, precise adjustments. Its head methodically turns from side to side, its amber eyes, sharp as glass beads, scanning the majestic mountains thousands of feet below with cool, focused attention. Below, an awe-inspiring panorama of rugged mountains stretches endlessly below, their peaks capped with snow, while deep green coniferous forests carpet the slopes like a thick carpet, and the exposed gray rock faces reveal rich textures in the light. A winding river, like a silver ribbon, weaves through a deep canyon, its waters reflecting dazzling light. Set against the backdrop of a vast blue sky, the entire image is imbued with a sense of tranquility and powerful energy, conveying the pristine, undisturbed beauty of nature and the solitary majesty of a top predator.", + "seed": 7846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_14/index_1938_seed_7846_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02937567327171564, + "vlm_logprob": -0.02937567327171564, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02937567327171564, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013885680586099625 + ], + "opposite_logprob_tokens": [ + -4.638885498046875 + ], + "logprob_reward": -0.013885680586099625, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013885680586099625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07808000594377518, + 0.09641341120004654, + 0.0433092899620533, + 0.043488286435604095, + 0.02847212366759777, + 0.06688692420721054, + 0.43942761421203613, + 0.0572800487279892, + 0.07065282762050629, + 0.07598944008350372 + ], + "credit_map": [ + 0.017809711396694183, + 0.0057689957320690155, + 0.011477004736661911, + 0.009302850812673569, + 0.005037042777985334, + 0.004895596764981747, + 0.014335328713059425, + 0.00945348385721445, + 0.022586097940802574, + 0.006660950370132923, + 0.015599608421325684, + 0.007429968565702438, + 0.005191716831177473, + 0.006072713993489742, + 0.015377582050859928, + 0.017494775354862213, + 0.004374712705612183, + 0.003337488742545247, + 0.008330674842000008, + 0.006894211750477552, + 0.003838061122223735, + 0.0031747017055749893, + 0.008386710658669472, + 0.004972732625901699, + 0.0035582594573497772, + 0.002766910707578063, + 0.008514650166034698, + 0.004498564638197422, + 0.00419471925124526, + 0.005311538465321064, + 0.005484752357006073, + 0.009158892557024956, + 0.003461360465735197, + 0.00270087830722332, + 0.004282430745661259, + 0.0036273745354264975, + 0.004380799364298582, + 0.002211690880358219, + 0.003585173049941659, + 0.004222417250275612, + 0.005702561233192682, + 0.006452187895774841, + 0.021171296015381813, + 0.009448854252696037, + 0.00639182748273015, + 0.003628816455602646, + 0.006349662318825722, + 0.007741718553006649, + 0.08732922375202179, + 0.11027027666568756, + 0.04516149312257767, + 0.05920511484146118, + 0.04273183271288872, + 0.02036381885409355, + 0.021798621863126755, + 0.052567269653081894, + 0.005756046157330275, + 0.005341377574950457, + 0.010641827248036861, + 0.00837110448628664, + 0.005721543449908495, + 0.007193929050117731, + 0.007832086645066738, + 0.006422135047614574, + 0.003927462734282017, + 0.005420773755759001, + 0.01275351457297802, + 0.008564140647649765, + 0.009910979308187962, + 0.011764710769057274, + 0.010437622666358948, + 0.007873624563217163, + 0.004204837139695883, + 0.006483318749815226, + 0.01080064196139574, + 0.010050670243799686, + 0.006996480282396078, + 0.012429594062268734, + 0.013919404707849026, + 0.011104496195912361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.013885680586099625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0142061123624444 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.0142061123624444, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0142061123624444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10686305165290833, + 0.19831643998622894, + 0.0836760401725769, + 0.07601221650838852, + 0.03508811071515083, + 0.09340907633304596, + 0.09348537772893906, + 0.0914410725235939, + 0.14573341608047485, + 0.07597518712282181 + ], + "credit_map": [ + 0.014950616285204887, + 0.008954296819865704, + 0.01638355851173401, + 0.013829373754560947, + 0.007049938663840294, + 0.006407647393643856, + 0.019099697470664978, + 0.020187929272651672, + 0.050785038620233536, + 0.014539756812155247, + 0.03495176509022713, + 0.01817398890852928, + 0.009848326444625854, + 0.01305421907454729, + 0.029990358278155327, + 0.026973005384206772, + 0.007955044507980347, + 0.0052460781298577785, + 0.016544505953788757, + 0.009824661538004875, + 0.006740237586200237, + 0.009334307163953781, + 0.01474402192980051, + 0.013287186622619629, + 0.007131928578019142, + 0.006022011395543814, + 0.011511556804180145, + 0.00924795214086771, + 0.010371477343142033, + 0.007503208238631487, + 0.01139958668500185, + 0.012824498116970062, + 0.004295440390706062, + 0.003067422192543745, + 0.006145693361759186, + 0.003217180259525776, + 0.003041792195290327, + 0.0029886451084166765, + 0.003332202322781086, + 0.008999738842248917, + 0.008438105694949627, + 0.0082756532356143, + 0.02954632230103016, + 0.01381970290094614, + 0.008722340688109398, + 0.006203396711498499, + 0.008929336443543434, + 0.009474227204918861, + 0.012049799785017967, + 0.017513317987322807, + 0.013450289145112038, + 0.014405034482479095, + 0.009694602340459824, + 0.006536457687616348, + 0.008478764444589615, + 0.011357123032212257, + 0.0082711111754179, + 0.010349154472351074, + 0.014785898849368095, + 0.011911415494978428, + 0.011930780485272408, + 0.011305312626063824, + 0.013123405165970325, + 0.009764003567397594, + 0.010374757461249828, + 0.011865547858178616, + 0.024248914793133736, + 0.01867850124835968, + 0.018201198428869247, + 0.02254440449178219, + 0.022423742339015007, + 0.017396356910467148, + 0.004848192445933819, + 0.006324951071292162, + 0.0097386809065938, + 0.010617156513035297, + 0.007128594443202019, + 0.012252447195351124, + 0.013711064122617245, + 0.011354110203683376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0142061123624444 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030385106801987 + ], + "opposite_logprob_tokens": [ + -2.350303888320923 + ], + "logprob_reward": -0.10030385106801987, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030385106801987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12539488077163696, + 0.11918646097183228, + 0.09211031347513199, + 0.05808277800679207, + 0.039901357144117355, + 0.08205612003803253, + 0.25502774119377136, + 0.057103127241134644, + 0.0941254273056984, + 0.07701176404953003 + ], + "credit_map": [ + 0.025267919525504112, + 0.010027427226305008, + 0.020411359146237373, + 0.014789948239922523, + 0.008051158860325813, + 0.006702445447444916, + 0.018573297187685966, + 0.021571330726146698, + 0.027253882959485054, + 0.008917439728975296, + 0.018826685845851898, + 0.010003176517784595, + 0.007156872656196356, + 0.010280769318342209, + 0.018285702913999557, + 0.018461937084794044, + 0.013043250888586044, + 0.007923636585474014, + 0.01601584441959858, + 0.011479310691356659, + 0.0058587659150362015, + 0.00615311972796917, + 0.01662369817495346, + 0.015012690797448158, + 0.005777609534561634, + 0.004386323969811201, + 0.011556864716112614, + 0.0064207338728010654, + 0.0055092996917665005, + 0.005924913100898266, + 0.008454236201941967, + 0.010052797384560108, + 0.005021295975893736, + 0.0035985803697258234, + 0.007151189725846052, + 0.003571871668100357, + 0.0038349241949617863, + 0.0028471597470343113, + 0.003543245606124401, + 0.010333091951906681, + 0.01355627179145813, + 0.007116944529116154, + 0.011548837646842003, + 0.011906532570719719, + 0.01171145774424076, + 0.005849367473274469, + 0.010108742862939835, + 0.010257967747747898, + 0.0484781414270401, + 0.06317202001810074, + 0.026084501296281815, + 0.037456415593624115, + 0.023697223514318466, + 0.01144903339445591, + 0.01372942142188549, + 0.030960997566580772, + 0.004666855093091726, + 0.005199403502047062, + 0.010768231935799122, + 0.007219117600470781, + 0.006710752844810486, + 0.006670170929282904, + 0.009605429135262966, + 0.006263167131692171, + 0.0058428021147847176, + 0.008294142782688141, + 0.013261227868497372, + 0.012578054331243038, + 0.010336401872336864, + 0.017807062715291977, + 0.013873199932277203, + 0.012132528237998486, + 0.004922138061374426, + 0.007833201438188553, + 0.012110681273043156, + 0.010616469196975231, + 0.009258260019123554, + 0.008097032085061073, + 0.013323291204869747, + 0.010850690305233002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.10030385106801987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539003044366837 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.012539003044366837, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539003044366837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08186285942792892, + 0.372026652097702, + 0.023779576644301414, + 0.014559624716639519, + 0.011715324595570564, + 0.05500168725848198, + 0.08080247044563293, + 0.17001385986804962, + 0.13674375414848328, + 0.053494155406951904 + ], + "credit_map": [ + 0.012523711659014225, + 0.005360393784940243, + 0.01569501683115959, + 0.008608751930296421, + 0.006328374147415161, + 0.005671148654073477, + 0.012885907664895058, + 0.014789566397666931, + 0.0989682525396347, + 0.023205552250146866, + 0.06309345364570618, + 0.02697516232728958, + 0.018816065043210983, + 0.024101603776216507, + 0.0526861809194088, + 0.0641804113984108, + 0.002722120378166437, + 0.0021410658955574036, + 0.004014944192022085, + 0.0035484095569700003, + 0.0021897456608712673, + 0.0014583716401830316, + 0.004898070823401213, + 0.0028068518731743097, + 0.0011898910161107779, + 0.0011028078151866794, + 0.0022614237386733294, + 0.0018273598980158567, + 0.0017365613020956516, + 0.001628876430913806, + 0.0024945358745753765, + 0.0023181710857897997, + 0.001285900711081922, + 0.0010961031075567007, + 0.002470454666763544, + 0.0009448091150261462, + 0.001114753889851272, + 0.0009187197429127991, + 0.0010642609558999538, + 0.002820324618369341, + 0.0056685274466872215, + 0.004986871499568224, + 0.019081389531493187, + 0.007580698002129793, + 0.003803956788033247, + 0.0039041712880134583, + 0.0042513711377978325, + 0.005724714137613773, + 0.014144807122647762, + 0.01924893818795681, + 0.008180911652743816, + 0.011321895755827427, + 0.007049936801195145, + 0.005259617231786251, + 0.005780298728495836, + 0.009816081263124943, + 0.019557584077119827, + 0.021655907854437828, + 0.027727803215384483, + 0.026542771607637405, + 0.014158830977976322, + 0.01888522133231163, + 0.022309500724077225, + 0.01917625032365322, + 0.008094359189271927, + 0.01082680094987154, + 0.02211727946996689, + 0.01701143942773342, + 0.015857722610235214, + 0.025668665766716003, + 0.0203211959451437, + 0.0168463047593832, + 0.0034434401895850897, + 0.004087475594133139, + 0.008589363656938076, + 0.00803916435688734, + 0.004022844135761261, + 0.007687609177082777, + 0.010189290158450603, + 0.007434975355863571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.012539003044366837 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943719297647476 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943719297647476, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943719297647476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10168618708848953, + 0.042721226811409, + 0.04699350893497467, + 0.01601928099989891, + 0.024471072480082512, + 0.1752733290195465, + 0.33059805631637573, + 0.10335001349449158, + 0.1021934375166893, + 0.056693919003009796 + ], + "credit_map": [ + 0.01501557044684887, + 0.007337562739849091, + 0.017681680619716644, + 0.014389190822839737, + 0.008233051747083664, + 0.006221410818397999, + 0.01711338944733143, + 0.015694323927164078, + 0.00666672270745039, + 0.0031138963531702757, + 0.007976914756000042, + 0.004854194819927216, + 0.002485740464180708, + 0.0023058990482240915, + 0.006337048951536417, + 0.008980813436210155, + 0.0050071305595338345, + 0.004101485479623079, + 0.007873840630054474, + 0.006141177378594875, + 0.004028733354061842, + 0.002984049031510949, + 0.008731735870242119, + 0.008125356398522854, + 0.0015394249930977821, + 0.0014844373799860477, + 0.0028820252045989037, + 0.00200264947488904, + 0.0020384069066494703, + 0.0014779212651774287, + 0.0023884656839072704, + 0.002205949043855071, + 0.0034505154471844435, + 0.0028318166732788086, + 0.0048800744116306305, + 0.0026111723855137825, + 0.00341108744032681, + 0.0017992120701819658, + 0.0026423917151987553, + 0.0028448039665818214, + 0.016847815364599228, + 0.015358136966824532, + 0.059042736887931824, + 0.02414851263165474, + 0.013107490725815296, + 0.01291628461331129, + 0.014491165988147259, + 0.019361192360520363, + 0.06374862790107727, + 0.08058410882949829, + 0.03419851139187813, + 0.043080803006887436, + 0.03589221462607384, + 0.014786948449909687, + 0.01803334802389145, + 0.04027349874377251, + 0.009269949048757553, + 0.01153754536062479, + 0.0166843943297863, + 0.01578412391245365, + 0.010432097129523754, + 0.012415271252393723, + 0.015415027737617493, + 0.011811602860689163, + 0.009098879992961884, + 0.00847835186868906, + 0.015807902440428734, + 0.01070715207606554, + 0.012888574041426182, + 0.013316229917109013, + 0.018964923918247223, + 0.012931418605148792, + 0.005084704142063856, + 0.003874225774779916, + 0.007604831829667091, + 0.007973670959472656, + 0.0060458616353571415, + 0.007301257457584143, + 0.010400993749499321, + 0.00840836763381958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.005943719297647476 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle aerial shot gracefully follows and slowly circles a majestic eagle, its massive, dark brown wings fully extended, effortlessly riding the high-altitude thermals. In the clear, high-altitude sunlight, the edges of each feather are rimmed with a golden hue, with the signature golden-brown feathers on the nape of its neck being particularly striking. The eagle maintains near-perfect stability, with only the wingtip feathers making minute, precise adjustments. Its head methodically turns from side to side, its amber eyes, sharp as glass beads, scanning the majestic mountains thousands of feet below with cool, focused attention. Below, an awe-inspiring panorama of rugged mountains stretches endlessly below, their peaks capped with snow, while deep green coniferous forests carpet the slopes like a thick carpet, and the exposed gray rock faces reveal rich textures in the light. A winding river, like a silver ribbon, weaves through a deep canyon, its waters reflecting dazzling light. Set against the backdrop of a vast blue sky, the entire image is imbued with a sense of tranquility and powerful energy, conveying the pristine, undisturbed beauty of nature and the solitary majesty of a top predator.", + "seed": 7847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_14/index_1938_seed_7847_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020314439851790666, + "vlm_logprob": -0.020314439851790666, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.020314439851790666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021502582356333733 + ], + "opposite_logprob_tokens": [ + -3.8965024948120117 + ], + "logprob_reward": -0.021502582356333733, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021502582356333733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3929249346256256, + 0.15869317948818207, + 0.05302329361438751, + 0.048489078879356384, + 0.050917696207761765, + 0.026094816625118256, + 0.05468789115548134, + 0.03691549599170685, + 0.09552415460348129, + 0.08272946625947952 + ], + "credit_map": [ + 0.04077179729938507, + 0.027820244431495667, + 0.08066024631261826, + 0.04099072143435478, + 0.11420875787734985, + 0.03960612416267395, + 0.022914081811904907, + 0.025952951982617378, + 0.010669163428246975, + 0.01147216372191906, + 0.01697695069015026, + 0.025962742045521736, + 0.013810756616294384, + 0.023294107988476753, + 0.03892914950847626, + 0.017578156664967537, + 0.003943334799259901, + 0.005271286703646183, + 0.011245913803577423, + 0.011745362542569637, + 0.0036552988458424807, + 0.003500262275338173, + 0.004774147178977728, + 0.008887686766684055, + 0.004452068358659744, + 0.004999896045774221, + 0.011307412758469582, + 0.004290565848350525, + 0.0076760463416576385, + 0.00304380152374506, + 0.004195101093500853, + 0.008524182252585888, + 0.005276829935610294, + 0.010319787077605724, + 0.009087426587939262, + 0.006194956600666046, + 0.005190721247345209, + 0.002717148046940565, + 0.004448096267879009, + 0.007682729512453079, + 0.0026545715518295765, + 0.0024539052974432707, + 0.005274113267660141, + 0.0034051556140184402, + 0.0021545439958572388, + 0.0026699595618993044, + 0.0033882837742567062, + 0.004094282630831003, + 0.007987314835190773, + 0.0074018193408846855, + 0.011392896063625813, + 0.005662957206368446, + 0.004419474396854639, + 0.00588261429220438, + 0.0032465385738760233, + 0.00869427714496851, + 0.007166932802647352, + 0.004291859455406666, + 0.005045139696449041, + 0.002975126728415489, + 0.0018128235824406147, + 0.0035849192645400763, + 0.003169682575389743, + 0.008869010023772717, + 0.009410102851688862, + 0.011068923398852348, + 0.013174550607800484, + 0.008922088891267776, + 0.010342644527554512, + 0.014074339531362057, + 0.013403710909187794, + 0.015127795748412609, + 0.008288554847240448, + 0.014577047899365425, + 0.013455577194690704, + 0.00958903506398201, + 0.00943358801305294, + 0.01218949630856514, + 0.006626571994274855, + 0.008569599129259586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.021502582356333733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588846877217293 + ], + "opposite_logprob_tokens": [ + -3.8955888748168945 + ], + "logprob_reward": -0.020588846877217293, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588846877217293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4335196018218994, + 0.09193584322929382, + 0.05800943076610565, + 0.028527982532978058, + 0.07394067198038101, + 0.04970531165599823, + 0.03997444733977318, + 0.05951152369379997, + 0.10232821851968765, + 0.06254693865776062 + ], + "credit_map": [ + 0.044171079993247986, + 0.030899005010724068, + 0.08036690205335617, + 0.04150853678584099, + 0.13362181186676025, + 0.048326727002859116, + 0.026047460734844208, + 0.028578098863363266, + 0.007066349498927593, + 0.0067757489159703255, + 0.012263043783605099, + 0.011771534569561481, + 0.008720194920897484, + 0.014588198624551296, + 0.01929548755288124, + 0.011455289088189602, + 0.003535269759595394, + 0.004665716551244259, + 0.012598597444593906, + 0.012163630686700344, + 0.004650379065424204, + 0.0037838465068489313, + 0.005576560273766518, + 0.011035427451133728, + 0.002583933062851429, + 0.0030965604819357395, + 0.0061400821432471275, + 0.0033584684133529663, + 0.0034723745193332434, + 0.0023298098240047693, + 0.00347536732442677, + 0.004071388393640518, + 0.006421074736863375, + 0.008585957810282707, + 0.017046790570020676, + 0.010443409904837608, + 0.006057392340153456, + 0.005337362177670002, + 0.00874970480799675, + 0.01129897776991129, + 0.0034603578969836235, + 0.0032521404791623354, + 0.007466915994882584, + 0.009951610118150711, + 0.0032389690168201923, + 0.005702993832528591, + 0.007288838271051645, + 0.00934348814189434, + 0.0042650578543543816, + 0.005703982897102833, + 0.005287122447043657, + 0.005252157803624868, + 0.003768615424633026, + 0.006097245961427689, + 0.003954037558287382, + 0.005646225530654192, + 0.011001307517290115, + 0.005894581321626902, + 0.006974369753152132, + 0.004741985350847244, + 0.003917214460670948, + 0.008399895392358303, + 0.005946761928498745, + 0.012635414488613605, + 0.009397242218255997, + 0.011862069368362427, + 0.014288168400526047, + 0.009646986611187458, + 0.013415715657174587, + 0.01517673209309578, + 0.015362580306828022, + 0.013178722932934761, + 0.0055033243261277676, + 0.008973952382802963, + 0.008584589697420597, + 0.006410608999431133, + 0.00662122992798686, + 0.006756964139640331, + 0.010389728471636772, + 0.009306537918746471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.020588846877217293 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486576184630394 + ], + "opposite_logprob_tokens": [ + -3.0486576557159424 + ], + "logprob_reward": -0.0486576184630394, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486576184630394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40916895866394043, + 0.03602893650531769, + 0.029867829754948616, + 0.025804029777646065, + 0.07053621858358383, + 0.01865467056632042, + 0.041617993265390396, + 0.014689861796796322, + 0.3029084801673889, + 0.050723034888505936 + ], + "credit_map": [ + 0.04011054337024689, + 0.02912525087594986, + 0.08095147460699081, + 0.04223988205194473, + 0.12765251100063324, + 0.039809029549360275, + 0.024399133399128914, + 0.024881144985556602, + 0.0039687552489340305, + 0.0032494093757122755, + 0.005267247557640076, + 0.006892841309309006, + 0.002552110468968749, + 0.00305883027613163, + 0.006853525526821613, + 0.0041862172074615955, + 0.003374637570232153, + 0.0025457951705902815, + 0.005820529069751501, + 0.005636598914861679, + 0.002554470207542181, + 0.0023028026334941387, + 0.0025634164921939373, + 0.005069578532129526, + 0.0028857076540589333, + 0.0032949717715382576, + 0.004266100004315376, + 0.002587807597592473, + 0.005268667824566364, + 0.0018744771368801594, + 0.0017747656675055623, + 0.003851530374959111, + 0.007026569452136755, + 0.013243517838418484, + 0.01091102696955204, + 0.009983963333070278, + 0.005814173258841038, + 0.005363982170820236, + 0.008112532086670399, + 0.010080455802381039, + 0.0022456857841461897, + 0.001777344150468707, + 0.0022818122524768114, + 0.0028727585449814796, + 0.0012056493433192372, + 0.001689799246378243, + 0.0027507073245942593, + 0.0038309143856167793, + 0.004864847287535667, + 0.006662226282060146, + 0.004723973106592894, + 0.006792688742280006, + 0.003943250980228186, + 0.0053957561030983925, + 0.0037064277566969395, + 0.005528824869543314, + 0.0020674800034612417, + 0.0023496150970458984, + 0.0025434952694922686, + 0.001610602717846632, + 0.0010741079458966851, + 0.0015853552613407373, + 0.0012403196888044477, + 0.002218885812908411, + 0.031501684337854385, + 0.02999761700630188, + 0.04474251717329025, + 0.017504535615444183, + 0.020982766523957253, + 0.07523791491985321, + 0.023373892530798912, + 0.05956755205988884, + 0.004524723626673222, + 0.004484846722334623, + 0.009102365002036095, + 0.006404097191989422, + 0.006868236232548952, + 0.005134685896337032, + 0.00901834201067686, + 0.005185734014958143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0486576184630394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732518319040537 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732518319040537, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732518319040537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4894815981388092, + 0.06727716326713562, + 0.03169490769505501, + 0.09148833900690079, + 0.14152154326438904, + 0.0336855873465538, + 0.02053394541144371, + 0.0359085388481617, + 0.06431593000888824, + 0.0240924172103405 + ], + "credit_map": [ + 0.04884469509124756, + 0.03840659186244011, + 0.07588743418455124, + 0.04010937735438347, + 0.15108951926231384, + 0.06392199546098709, + 0.029734961688518524, + 0.04148704558610916, + 0.0060736630111932755, + 0.005364404059946537, + 0.008356847800314426, + 0.010449059307575226, + 0.006461513694375753, + 0.01197414007037878, + 0.00961286947131157, + 0.008984662592411041, + 0.00263065192848444, + 0.003034338355064392, + 0.0063370754942297935, + 0.005937657784670591, + 0.0031451499089598656, + 0.0025973678566515446, + 0.0031405508052557707, + 0.004872115794569254, + 0.007538548670709133, + 0.010306610725820065, + 0.014185902662575245, + 0.011096904054284096, + 0.015254846774041653, + 0.0060139806009829044, + 0.0108342245221138, + 0.01625731959939003, + 0.01683238334953785, + 0.034973785281181335, + 0.020491942763328552, + 0.012873590923845768, + 0.01466448325663805, + 0.007360766641795635, + 0.010112103074789047, + 0.02421247959136963, + 0.0037580847274512053, + 0.0031283036805689335, + 0.0047555482015013695, + 0.009328192099928856, + 0.0020365563686937094, + 0.001742758322507143, + 0.0027126804925501347, + 0.006223461590707302, + 0.002272288780659437, + 0.003912272397428751, + 0.004352355841547251, + 0.0023372049909085035, + 0.0016525402897968888, + 0.0015797932865098119, + 0.0014264865312725306, + 0.0030010021291673183, + 0.005000030621886253, + 0.006351514719426632, + 0.006089217960834503, + 0.003425477771088481, + 0.0026492688339203596, + 0.003208078444004059, + 0.0026393134612590075, + 0.006545641925185919, + 0.005221461411565542, + 0.006186858285218477, + 0.008688880130648613, + 0.005937235429883003, + 0.005391987971961498, + 0.01494090911000967, + 0.00630329130217433, + 0.011645308695733547, + 0.003057706868276, + 0.002620173618197441, + 0.003519588615745306, + 0.0025598069187253714, + 0.0029273040127009153, + 0.003972623497247696, + 0.0018910868093371391, + 0.0035441261716187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.006732518319040537 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040906332433223724 + ], + "opposite_logprob_tokens": [ + -5.504090785980225 + ], + "logprob_reward": -0.0040906332433223724, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040906332433223724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39242514967918396, + 0.1487356424331665, + 0.09138957411050797, + 0.06309147924184799, + 0.0656469389796257, + 0.03686884790658951, + 0.03648659959435463, + 0.032759156078100204, + 0.09460406005382538, + 0.03799253702163696 + ], + "credit_map": [ + 0.036735158413648605, + 0.026442889124155045, + 0.09711072593927383, + 0.05127934738993645, + 0.09276783466339111, + 0.032178618013858795, + 0.025825874879956245, + 0.03008471429347992, + 0.011009925976395607, + 0.009092766791582108, + 0.014998098835349083, + 0.017781388014554977, + 0.014332146383821964, + 0.026786375790834427, + 0.02800976112484932, + 0.026725171133875847, + 0.010849474929273129, + 0.00752905011177063, + 0.022622182965278625, + 0.015411249361932278, + 0.007458211854100227, + 0.004915889352560043, + 0.007682614494115114, + 0.014920898713171482, + 0.007900748401880264, + 0.011763517744839191, + 0.006539664696902037, + 0.006424038205295801, + 0.015405165031552315, + 0.00422437209635973, + 0.0043105254881083965, + 0.006523445714265108, + 0.007968536578118801, + 0.015939047560095787, + 0.008531207218766212, + 0.00506423506885767, + 0.009504065848886967, + 0.003773051081225276, + 0.0062807961367070675, + 0.008585994131863117, + 0.002578876679763198, + 0.0027907199691981077, + 0.007030799053609371, + 0.006518486887216568, + 0.0027328876312822104, + 0.005027825944125652, + 0.004696344491094351, + 0.0054929074831306934, + 0.005945852026343346, + 0.0054269530810415745, + 0.006976508069783449, + 0.003209275659173727, + 0.0025475157890468836, + 0.0038872279692441225, + 0.0035559888929128647, + 0.004937279503792524, + 0.003439274150878191, + 0.00533892959356308, + 0.007846345193684101, + 0.0031431762035936117, + 0.002227774588391185, + 0.0042120954021811485, + 0.0024492309894412756, + 0.004102329723536968, + 0.011400965973734856, + 0.009931549429893494, + 0.012390987947583199, + 0.00921747088432312, + 0.011300520971417427, + 0.014137929305434227, + 0.013099984265863895, + 0.013124643824994564, + 0.003443213179707527, + 0.0045035965740680695, + 0.005255363881587982, + 0.0047856601886451244, + 0.004085205029696226, + 0.005279684439301491, + 0.005338283721357584, + 0.005301528610289097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0040906332433223724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, high-angle aerial shot gracefully follows and slowly circles a majestic eagle, its massive, dark brown wings fully extended, effortlessly riding the high-altitude thermals. In the clear, high-altitude sunlight, the edges of each feather are rimmed with a golden hue, with the signature golden-brown feathers on the nape of its neck being particularly striking. The eagle maintains near-perfect stability, with only the wingtip feathers making minute, precise adjustments. Its head methodically turns from side to side, its amber eyes, sharp as glass beads, scanning the majestic mountains thousands of feet below with cool, focused attention. Below, an awe-inspiring panorama of rugged mountains stretches endlessly below, their peaks capped with snow, while deep green coniferous forests carpet the slopes like a thick carpet, and the exposed gray rock faces reveal rich textures in the light. A winding river, like a silver ribbon, weaves through a deep canyon, its waters reflecting dazzling light. Set against the backdrop of a vast blue sky, the entire image is imbued with a sense of tranquility and powerful energy, conveying the pristine, undisturbed beauty of nature and the solitary majesty of a top predator.", + "seed": 7848, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_14/index_1938_seed_7848_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015643206099048255, + "vlm_logprob": -0.015643206099048255, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.015643206099048255, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005609485320746899 + ], + "opposite_logprob_tokens": [ + -5.505609512329102 + ], + "logprob_reward": -0.005609485320746899, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005609485320746899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27654191851615906, + 0.24718117713928223, + 0.05795151740312576, + 0.06965428590774536, + 0.04219089075922966, + 0.08433020114898682, + 0.055745989084243774, + 0.0972607210278511, + 0.037655267864465714, + 0.031487930566072464 + ], + "credit_map": [ + 0.021184122189879417, + 0.016614411026239395, + 0.03185415640473366, + 0.03510710597038269, + 0.019178828224539757, + 0.029618283733725548, + 0.0733875185251236, + 0.04959755763411522, + 0.023869488388299942, + 0.01429012417793274, + 0.03206658363342285, + 0.015573378652334213, + 0.012668220326304436, + 0.02119688130915165, + 0.04464372992515564, + 0.08287280052900314, + 0.004452715627849102, + 0.0033120461739599705, + 0.00777915446087718, + 0.007108916528522968, + 0.004048047121614218, + 0.0037847214844077826, + 0.016377052292227745, + 0.011088875122368336, + 0.004308369942009449, + 0.005308335646986961, + 0.00951852835714817, + 0.01589890755712986, + 0.007432907819747925, + 0.005963750649243593, + 0.010172974318265915, + 0.011050528846681118, + 0.003696748288348317, + 0.0043067303486168385, + 0.007064071949571371, + 0.006450130138546228, + 0.005704058334231377, + 0.003757206490263343, + 0.007262132596224546, + 0.0039498223923146725, + 0.010041571222245693, + 0.007806543726474047, + 0.013510126620531082, + 0.015170122496783733, + 0.008500111289322376, + 0.010541923344135284, + 0.007609065156430006, + 0.01115075871348381, + 0.005277003161609173, + 0.004908172879368067, + 0.007637289818376303, + 0.005801561288535595, + 0.00901001039892435, + 0.0119222616776824, + 0.004066716879606247, + 0.00712298508733511, + 0.00703698443248868, + 0.0033890868071466684, + 0.02503490261733532, + 0.009315830655395985, + 0.003869708627462387, + 0.008625025860965252, + 0.007711521815508604, + 0.032277680933475494, + 0.004858321510255337, + 0.0043260459788143635, + 0.005369867663830519, + 0.004621499218046665, + 0.0036982924211770296, + 0.00575985386967659, + 0.005063926801085472, + 0.00395746948197484, + 0.003151928773149848, + 0.002751524792984128, + 0.00382565101608634, + 0.0032583947759121656, + 0.00412558251991868, + 0.006603860296308994, + 0.0034903129562735558, + 0.004280681721866131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.005609485320746899 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083937250077724 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011083937250077724, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083937250077724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1947624683380127, + 0.0908142626285553, + 0.1343412697315216, + 0.07014194875955582, + 0.07188662141561508, + 0.07180511951446533, + 0.03916531056165695, + 0.17807094752788544, + 0.05035821348428726, + 0.09865391254425049 + ], + "credit_map": [ + 0.010420256294310093, + 0.008735992014408112, + 0.027362024411559105, + 0.017345860600471497, + 0.013118095695972443, + 0.023845594376325607, + 0.05905337631702423, + 0.03488126024603844, + 0.007320464123040438, + 0.005641087889671326, + 0.012090446427464485, + 0.006980712525546551, + 0.00516637833788991, + 0.008516631089150906, + 0.01550863403826952, + 0.029589902609586716, + 0.005936813540756702, + 0.007047160062938929, + 0.01729177124798298, + 0.015461900271475315, + 0.007262587547302246, + 0.007777591235935688, + 0.027325348928570747, + 0.04623810201883316, + 0.00689716124907136, + 0.004176085814833641, + 0.00849232915788889, + 0.011394420638680458, + 0.0062943133525550365, + 0.006115068681538105, + 0.01699173077940941, + 0.009780834428966045, + 0.005016898270696402, + 0.005163442809134722, + 0.014562723226845264, + 0.01134136226028204, + 0.007587232626974583, + 0.006834407337009907, + 0.010504269041121006, + 0.01087628398090601, + 0.008004263043403625, + 0.004647672176361084, + 0.01230647973716259, + 0.01246042549610138, + 0.00705622648820281, + 0.013605182059109211, + 0.005828215274959803, + 0.007896657101809978, + 0.0037014808040112257, + 0.0033064810559153557, + 0.005669006612151861, + 0.004219175782054663, + 0.006052349228411913, + 0.008489417843520641, + 0.0030646128579974174, + 0.004662784282118082, + 0.012523235753178596, + 0.007431150879710913, + 0.042170193046331406, + 0.01669267565011978, + 0.008560347370803356, + 0.018897205591201782, + 0.013295067474246025, + 0.05850107595324516, + 0.005271912552416325, + 0.004275798797607422, + 0.006018146872520447, + 0.0062174745835363865, + 0.006969664711505175, + 0.009027178399264812, + 0.0075020138174295425, + 0.005076020490378141, + 0.006890891119837761, + 0.0077451602555811405, + 0.011576601304113865, + 0.011771893128752708, + 0.015187719836831093, + 0.015187772922217846, + 0.018434932455420494, + 0.011858940124511719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.011083937250077724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495012179017067 + ], + "opposite_logprob_tokens": [ + -2.92995023727417 + ], + "logprob_reward": -0.05495012179017067, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495012179017067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10802052915096283, + 0.06203485652804375, + 0.17491357028484344, + 0.20504961907863617, + 0.09742731600999832, + 0.06237645819783211, + 0.02726093679666519, + 0.13554419577121735, + 0.04250296205282211, + 0.08486955612897873 + ], + "credit_map": [ + 0.0052025397308170795, + 0.00842535775154829, + 0.01676737703382969, + 0.013248216360807419, + 0.008842355571687222, + 0.009183435700833797, + 0.021972104907035828, + 0.02437913417816162, + 0.004825871903449297, + 0.004121152684092522, + 0.009290856309235096, + 0.007217610254883766, + 0.0034501091577112675, + 0.006184520665556192, + 0.009534395299851894, + 0.017410343512892723, + 0.00718724587932229, + 0.008750736713409424, + 0.026267582550644875, + 0.017066119238734245, + 0.011182512156665325, + 0.009263850748538971, + 0.036035049706697464, + 0.059160489588975906, + 0.02275237627327442, + 0.012013744562864304, + 0.028678376227617264, + 0.028383536264300346, + 0.01797754503786564, + 0.012697672471404076, + 0.0558144636452198, + 0.026731904596090317, + 0.009278042241930962, + 0.006681655067950487, + 0.018744831904768944, + 0.010833640582859516, + 0.010626420378684998, + 0.010358855128288269, + 0.01294375117868185, + 0.017960110679268837, + 0.004815912805497646, + 0.005711188539862633, + 0.008779579773545265, + 0.009523958899080753, + 0.007355242036283016, + 0.009697514586150646, + 0.005568955559283495, + 0.01092410646378994, + 0.0031265001744031906, + 0.0019175097113475204, + 0.004021315835416317, + 0.004564439877867699, + 0.0032115092035382986, + 0.003060293383896351, + 0.002813776023685932, + 0.004545592702925205, + 0.00891224667429924, + 0.0048555596731603146, + 0.032338764518499374, + 0.013318478129804134, + 0.008024836890399456, + 0.012178550474345684, + 0.010725192725658417, + 0.045190565288066864, + 0.0036559980362653732, + 0.0036546620540320873, + 0.008020120672881603, + 0.007330411113798618, + 0.005078690592199564, + 0.006351429503411055, + 0.00433771638199687, + 0.004073936026543379, + 0.004953102674335241, + 0.006359755527228117, + 0.0108089754357934, + 0.01450465340167284, + 0.007001572288572788, + 0.016013620421290398, + 0.013857770711183548, + 0.011370099149644375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.05495012179017067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634472541511059 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634472541511059, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634472541511059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19117993116378784, + 0.1915951520204544, + 0.05606059357523918, + 0.1080419197678566, + 0.06482543796300888, + 0.03228644281625748, + 0.030130980536341667, + 0.10566269606351852, + 0.10036865621805191, + 0.11984822154045105 + ], + "credit_map": [ + 0.011498437263071537, + 0.01021991204470396, + 0.024022774770855904, + 0.019680645316839218, + 0.012643951922655106, + 0.022244645282626152, + 0.0568401962518692, + 0.03402937203645706, + 0.01803632266819477, + 0.01219425443559885, + 0.024880891665816307, + 0.015291610732674599, + 0.009966159239411354, + 0.01312815211713314, + 0.03496313840150833, + 0.06313462555408478, + 0.0037902281619608402, + 0.004318798892199993, + 0.009853018447756767, + 0.007618824020028114, + 0.007349932100623846, + 0.003423584159463644, + 0.011214381083846092, + 0.008491831831634045, + 0.010020526126027107, + 0.00788323488086462, + 0.020658578723669052, + 0.013045311905443668, + 0.01515812985599041, + 0.00907827913761139, + 0.017403215169906616, + 0.014794642105698586, + 0.009944407269358635, + 0.005206571891903877, + 0.012800060212612152, + 0.012916093692183495, + 0.004377765115350485, + 0.005277479998767376, + 0.006878752261400223, + 0.007424307055771351, + 0.003567138919606805, + 0.002892429241910577, + 0.00500526512041688, + 0.006365817505866289, + 0.003173308912664652, + 0.00387922883965075, + 0.00284120487049222, + 0.004562049172818661, + 0.003957075998187065, + 0.002122314181178808, + 0.005444291979074478, + 0.00393123971298337, + 0.0032334530260413885, + 0.0033401071559637785, + 0.002579108579084277, + 0.005523388274013996, + 0.0085173100233078, + 0.00500075239688158, + 0.026866910979151726, + 0.008320039138197899, + 0.003830259433016181, + 0.009424316696822643, + 0.008777433075010777, + 0.0349256694316864, + 0.007036251947283745, + 0.005782699212431908, + 0.01526463869959116, + 0.012687476351857185, + 0.012069093063473701, + 0.018888860940933228, + 0.016884494572877884, + 0.0117551414296031, + 0.0094855772331357, + 0.007178382482379675, + 0.010832348838448524, + 0.01174042746424675, + 0.01677430421113968, + 0.03052237071096897, + 0.01468407642096281, + 0.018630724400281906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.004634472541511059 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a large eagle with dark brown wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle have golden-brown feathers on the nape of its neck and golden-rimmed feather edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the eagle's head turning from side to side while its amber eyes scan the landscape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a high-altitude panorama featuring snow-capped mountains, green coniferous forests, and a winding silver river in a canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a sweeping high-angle aerial view that gracefully follows and slowly circles the eagle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001938013592734933 + ], + "opposite_logprob_tokens": [ + -6.2519378662109375 + ], + "logprob_reward": -0.001938013592734933, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001938013592734933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13712255656719208, + 0.1810353398323059, + 0.10989446192979813, + 0.1769886016845703, + 0.0748741403222084, + 0.03566921874880791, + 0.0396028570830822, + 0.07592200487852097, + 0.051521461457014084, + 0.11736933887004852 + ], + "credit_map": [ + 0.01016984786838293, + 0.010369242168962955, + 0.0161062590777874, + 0.022882236167788506, + 0.00996359158307314, + 0.01635553129017353, + 0.0259141493588686, + 0.025361696258187294, + 0.01371560338884592, + 0.009415681473910809, + 0.02523195743560791, + 0.011402403935790062, + 0.007837842218577862, + 0.013293487951159477, + 0.03469539061188698, + 0.06544298678636551, + 0.004505804739892483, + 0.003821329213678837, + 0.01668294332921505, + 0.015823369845747948, + 0.006033540703356266, + 0.005783938802778721, + 0.028853794559836388, + 0.028389740735292435, + 0.021299894899129868, + 0.010705890133976936, + 0.025235548615455627, + 0.02444688230752945, + 0.013441345654428005, + 0.010141799226403236, + 0.04780152067542076, + 0.023915721103549004, + 0.006762638222426176, + 0.006366584450006485, + 0.014362383633852005, + 0.013608420267701149, + 0.006959463004022837, + 0.006374547723680735, + 0.010622686706483364, + 0.009817413985729218, + 0.003353775478899479, + 0.0027745976112782955, + 0.005546643398702145, + 0.007462986744940281, + 0.0033645571675151587, + 0.003667659591883421, + 0.003453043522313237, + 0.006045956164598465, + 0.003306445898488164, + 0.0028690698090940714, + 0.0062651424668729305, + 0.006222968455404043, + 0.0043859221041202545, + 0.004561133217066526, + 0.004525972530245781, + 0.0074662030674517155, + 0.005475232843309641, + 0.004057104233652353, + 0.018067920580506325, + 0.006981766317039728, + 0.005689363460987806, + 0.0072240643203258514, + 0.008372814394533634, + 0.02005373314023018, + 0.007009819149971008, + 0.004847920034080744, + 0.007288177963346243, + 0.006298063322901726, + 0.005748386029154062, + 0.008581108413636684, + 0.0062027499079704285, + 0.005545234307646751, + 0.006660218350589275, + 0.009006320498883724, + 0.014981362968683243, + 0.020162541419267654, + 0.010430006310343742, + 0.02313695289194584, + 0.0173000767827034, + 0.01569186896085739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.001938013592734933 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_15/index_tensor([3776, 3776])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_15/index_tensor([3776, 3776])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0c49baefc43e91f832daf0c6aea9c37627898e3d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_15/index_tensor([3776, 3776])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "strikingly realistic slow-motion close-up, the camera held still, captures from eye level the moment when billowing white steam slowly rises from a traditional bamboo steamer. This warm, moist vapor gently expands and swirls in the air, delicately intertwining with the cool morning mist seeping in through the window, forming a hazy, dynamic curtain. Within this shrouded haze, the gentle profile of a middle-aged mother emerges, her raven hair pulled back in a simple bun. A few sweat-soaked strands cling to her cheeks, and a few grains of glistening rice, unnoticed by her busy schedule, are clearly visible at her temples. The soft morning light penetrates the mist, outlining a soft edge of light for her figure and making the few grains of rice slightly reflect light. Her focused and serene expression appears particularly moving in the steam that is sometimes thick and sometimes thin. The whole picture, with its blend of warm and cool tones, creates an extremely tranquil and nostalgic atmosphere full of life and deep emotions.", + "seed": 26265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_15/index_3776_seed_26265_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.23355071125552057, + "vlm_logprob": -0.23355071125552057, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.23355071125552057, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014293666929006577 + ], + "opposite_logprob_tokens": [ + -4.514293670654297 + ], + "logprob_reward": -0.014293666929006577, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014293666929006577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2691538333892822, + 0.09610418975353241, + 0.10990259051322937, + 0.10704846680164337, + 0.12223535031080246, + 0.043083325028419495, + 0.07119829952716827, + 0.06807044893503189, + 0.05639608949422836, + 0.05680738016963005 + ], + "credit_map": [ + 0.027735529467463493, + 0.040011726319789886, + 0.0409439355134964, + 0.035007502883672714, + 0.0291715357452631, + 0.03518325462937355, + 0.04033719003200531, + 0.020763177424669266, + 0.011786265298724174, + 0.011383973062038422, + 0.011210910975933075, + 0.01297851838171482, + 0.00749399047344923, + 0.01735493168234825, + 0.0147637277841568, + 0.009131872095167637, + 0.008031961508095264, + 0.011409007012844086, + 0.01397034339606762, + 0.010411140508949757, + 0.014270113781094551, + 0.02131216786801815, + 0.01699083298444748, + 0.013507022522389889, + 0.010144283063709736, + 0.011745198629796505, + 0.01140040997415781, + 0.007971027866005898, + 0.017544973641633987, + 0.013508922420442104, + 0.02098710834980011, + 0.013746540993452072, + 0.011365368962287903, + 0.014583670534193516, + 0.01914221979677677, + 0.02257700264453888, + 0.015255662612617016, + 0.013607801869511604, + 0.011812913231551647, + 0.013890722766518593, + 0.0038423947989940643, + 0.00624256394803524, + 0.008669842034578323, + 0.006970275659114122, + 0.004715594928711653, + 0.004147535189986229, + 0.003773447358980775, + 0.004721674136817455, + 0.0069840517826378345, + 0.007533589377999306, + 0.006934845354408026, + 0.011185560375452042, + 0.009492580778896809, + 0.014839774928987026, + 0.007835526950657368, + 0.006392375100404024, + 0.010947495698928833, + 0.010432131588459015, + 0.006149055436253548, + 0.00858874898403883, + 0.007747895549982786, + 0.009069599211215973, + 0.004987131338566542, + 0.010148388333618641, + 0.004801071714609861, + 0.008904125541448593, + 0.008276213891804218, + 0.01076512597501278, + 0.005429685581475496, + 0.006306876894086599, + 0.005495652090758085, + 0.006417338736355305, + 0.004921666346490383, + 0.00640550721436739, + 0.004110376816242933, + 0.006399655714631081, + 0.01123191136866808, + 0.011027233675122261, + 0.0047892979346215725, + 0.007921736687421799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.014293666929006577 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539298057556152 + ], + "opposite_logprob_tokens": [ + -0.4289298355579376 + ], + "logprob_reward": -1.0539298057556152, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539298057556152, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15376821160316467, + 0.12149786949157715, + 0.13596490025520325, + 0.08419466763734818, + 0.09705597162246704, + 0.15186816453933716, + 0.06645085662603378, + 0.04559779539704323, + 0.05715429037809372, + 0.08644722402095795 + ], + "credit_map": [ + 0.016943665221333504, + 0.026377862319350243, + 0.02164958417415619, + 0.021289611235260963, + 0.01826833188533783, + 0.02070862241089344, + 0.018520038574934006, + 0.010010494850575924, + 0.0168101005256176, + 0.013478392735123634, + 0.020978894084692, + 0.024662373587489128, + 0.007004146929830313, + 0.017455779016017914, + 0.013075054623186588, + 0.008033138699829578, + 0.012454034760594368, + 0.02092595025897026, + 0.012934993952512741, + 0.01554593164473772, + 0.021577510982751846, + 0.01950807496905327, + 0.022071298211812973, + 0.010947104543447495, + 0.005435924511402845, + 0.010949120856821537, + 0.0082329660654068, + 0.01022863294929266, + 0.01271263137459755, + 0.016553819179534912, + 0.011931536719202995, + 0.008150034584105015, + 0.005825139116495848, + 0.015065285377204418, + 0.012522774748504162, + 0.016709385439753532, + 0.012460947036743164, + 0.010905188508331776, + 0.011124241165816784, + 0.012443017214536667, + 0.011572804301977158, + 0.02032468095421791, + 0.03469102084636688, + 0.020655956119298935, + 0.017910290509462357, + 0.02007414400577545, + 0.011151239275932312, + 0.01548803597688675, + 0.006991409696638584, + 0.012973522767424583, + 0.005779446568340063, + 0.01209090743213892, + 0.007538709323853254, + 0.007126401644200087, + 0.007762840483337641, + 0.006187614519149065, + 0.003280844772234559, + 0.007369904313236475, + 0.004008902702480555, + 0.007000260055065155, + 0.00482166837900877, + 0.007557090371847153, + 0.004633347038179636, + 0.006925777532160282, + 0.004097682423889637, + 0.006881366949528456, + 0.006269003730267286, + 0.009904460050165653, + 0.006551310885697603, + 0.009979776106774807, + 0.005963423755019903, + 0.007507265079766512, + 0.007943976670503616, + 0.011035348288714886, + 0.007861229591071606, + 0.011498620733618736, + 0.014400490559637547, + 0.01642702706158161, + 0.009489425458014011, + 0.007791104260832071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -1.0539298057556152 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022025867365300655 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022025867365300655, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022025867365300655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19425223767757416, + 0.08375458419322968, + 0.23582790791988373, + 0.11108437925577164, + 0.07589590549468994, + 0.06619039177894592, + 0.054706841707229614, + 0.08911196887493134, + 0.04943832755088806, + 0.039737436920404434 + ], + "credit_map": [ + 0.021102508530020714, + 0.033348198980093, + 0.030057623982429504, + 0.027002649381756783, + 0.01865365542471409, + 0.02241692505776882, + 0.025645717978477478, + 0.016024960204958916, + 0.009447513148188591, + 0.009252186864614487, + 0.011725551448762417, + 0.012452107854187489, + 0.00717201828956604, + 0.013317858800292015, + 0.013568082824349403, + 0.006819261237978935, + 0.01764124259352684, + 0.02775534987449646, + 0.019676903262734413, + 0.02478358894586563, + 0.04497122764587402, + 0.04180004820227623, + 0.036402635276317596, + 0.02279692329466343, + 0.006547408644109964, + 0.01318071037530899, + 0.010571104474365711, + 0.010059008374810219, + 0.01902550645172596, + 0.014619451016187668, + 0.02776884101331234, + 0.009312337264418602, + 0.004594484344124794, + 0.009191102348268032, + 0.01151000801473856, + 0.011798059567809105, + 0.01589621789753437, + 0.008665834553539753, + 0.008588215336203575, + 0.005651983432471752, + 0.005638648755848408, + 0.008460428565740585, + 0.009363813325762749, + 0.01050665881484747, + 0.010236123576760292, + 0.01016235537827015, + 0.005753307603299618, + 0.006069055758416653, + 0.003533635288476944, + 0.012991946190595627, + 0.004721975885331631, + 0.006599282380193472, + 0.004549485631287098, + 0.008125578984618187, + 0.005897131748497486, + 0.008287801407277584, + 0.008084067143499851, + 0.014794304035604, + 0.006532387807965279, + 0.01074287761002779, + 0.006862543057650328, + 0.021226223558187485, + 0.009992842562496662, + 0.010876725427806377, + 0.004815843887627125, + 0.009374156594276428, + 0.005036594811826944, + 0.008679325692355633, + 0.005717234220355749, + 0.006101729813963175, + 0.005723875481635332, + 0.003989566583186388, + 0.003526895772665739, + 0.004845394287258387, + 0.003336446126922965, + 0.005116559565067291, + 0.005162306595593691, + 0.009475051425397396, + 0.0038771918043494225, + 0.004397591110318899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0022025867365300655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048660796135663986 + ], + "opposite_logprob_tokens": [ + -3.0486607551574707 + ], + "logprob_reward": -0.048660796135663986, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048660796135663986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1724015772342682, + 0.13160587847232819, + 0.1985156089067459, + 0.07646670937538147, + 0.08430826663970947, + 0.04605412483215332, + 0.08602628856897354, + 0.08622252941131592, + 0.07293373346328735, + 0.045465338975191116 + ], + "credit_map": [ + 0.018179798498749733, + 0.03162839636206627, + 0.026182066649198532, + 0.021708697080612183, + 0.019193237647414207, + 0.02057665027678013, + 0.018222149461507797, + 0.016710568219423294, + 0.016450121998786926, + 0.017330199480056763, + 0.02351003885269165, + 0.01555436011403799, + 0.01967417635023594, + 0.01852458529174328, + 0.011154444888234138, + 0.009407952427864075, + 0.01227281242609024, + 0.018545161932706833, + 0.022944815456867218, + 0.01197456568479538, + 0.028698882088065147, + 0.014324212446808815, + 0.07042600959539413, + 0.01932913064956665, + 0.005144130438566208, + 0.010302986949682236, + 0.008113184943795204, + 0.0075146411545574665, + 0.013106792233884335, + 0.009370245970785618, + 0.014761459082365036, + 0.008153262548148632, + 0.007532821968197823, + 0.008824408054351807, + 0.015170851722359657, + 0.011626705527305603, + 0.015392533503472805, + 0.009322799742221832, + 0.007937873713672161, + 0.008500264957547188, + 0.0036178738810122013, + 0.005740545690059662, + 0.007750255521386862, + 0.00656643882393837, + 0.005968600977212191, + 0.0078086573630571365, + 0.00382662913762033, + 0.004775120411068201, + 0.0055432976223528385, + 0.01205950602889061, + 0.007566309999674559, + 0.01777372509241104, + 0.01356633473187685, + 0.012090410105884075, + 0.011536171659827232, + 0.005890522617846727, + 0.009167597629129887, + 0.01218761969357729, + 0.007446561940014362, + 0.012057771906256676, + 0.006993342190980911, + 0.02003609761595726, + 0.007449586410075426, + 0.010883940383791924, + 0.008966326713562012, + 0.011298087425529957, + 0.007195988204330206, + 0.011724867857992649, + 0.00899738259613514, + 0.006940193474292755, + 0.011682398617267609, + 0.006128482520580292, + 0.004351999145001173, + 0.00616711750626564, + 0.003477072808891535, + 0.007037445902824402, + 0.006157300900667906, + 0.006360144354403019, + 0.003988129086792469, + 0.007926125079393387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.048660796135663986 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866670072078705 + ], + "opposite_logprob_tokens": [ + -3.0486667156219482 + ], + "logprob_reward": -0.04866670072078705, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866670072078705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19424286484718323, + 0.1541580706834793, + 0.12223391234874725, + 0.06893715262413025, + 0.08556842058897018, + 0.043029434978961945, + 0.09915183484554291, + 0.0681908130645752, + 0.041590455919504166, + 0.12289708107709885 + ], + "credit_map": [ + 0.02120923437178135, + 0.03212067484855652, + 0.028867216780781746, + 0.02975982055068016, + 0.01620928756892681, + 0.022633371874690056, + 0.029004743322730064, + 0.0144385090097785, + 0.016293808817863464, + 0.01306107547134161, + 0.035685501992702484, + 0.016937239095568657, + 0.0265936441719532, + 0.02165481634438038, + 0.016601094976067543, + 0.007330895867198706, + 0.011567974463105202, + 0.01604984700679779, + 0.014139669016003609, + 0.012121469713747501, + 0.021365484222769737, + 0.020314229652285576, + 0.014211754314601421, + 0.01246348675340414, + 0.004734940826892853, + 0.005811711307615042, + 0.006288536824285984, + 0.008253669366240501, + 0.010253245010972023, + 0.015955612063407898, + 0.010265083983540535, + 0.007374353241175413, + 0.0055051567032933235, + 0.009951275773346424, + 0.012159994803369045, + 0.015142532996833324, + 0.0140814995393157, + 0.01281656976789236, + 0.009285638108849525, + 0.006625758018344641, + 0.003990881610661745, + 0.004815347958356142, + 0.004712644498795271, + 0.005872425157576799, + 0.007132934406399727, + 0.007661798503249884, + 0.004346410743892193, + 0.004496992565691471, + 0.008306052535772324, + 0.03032105602324009, + 0.00679194089025259, + 0.0080718407407403, + 0.012517406605184078, + 0.007656933274120092, + 0.00848865695297718, + 0.01699795015156269, + 0.006078050471842289, + 0.008791488595306873, + 0.004865158814936876, + 0.008517388254404068, + 0.005539598874747753, + 0.017709646373987198, + 0.007421336602419615, + 0.0092681385576725, + 0.004149146843701601, + 0.005431548226624727, + 0.004074358846992254, + 0.007480270229279995, + 0.004411978181451559, + 0.006524764932692051, + 0.004997840616852045, + 0.004520552232861519, + 0.010919254273176193, + 0.01631741039454937, + 0.0099445516243577, + 0.01967666484415531, + 0.021685097366571426, + 0.018464569002389908, + 0.01484735868871212, + 0.01104216929525137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.04866670072078705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly realistic slow-motion close-up, the camera held still, captures from eye level the moment when billowing white steam slowly rises from a traditional bamboo steamer. This warm, moist vapor gently expands and swirls in the air, delicately intertwining with the cool morning mist seeping in through the window, forming a hazy, dynamic curtain. Within this shrouded haze, the gentle profile of a middle-aged mother emerges, her raven hair pulled back in a simple bun. A few sweat-soaked strands cling to her cheeks, and a few grains of glistening rice, unnoticed by her busy schedule, are clearly visible at her temples. The soft morning light penetrates the mist, outlining a soft edge of light for her figure and making the few grains of rice slightly reflect light. Her focused and serene expression appears particularly moving in the steam that is sometimes thick and sometimes thin. The whole picture, with its blend of warm and cool tones, creates an extremely tranquil and nostalgic atmosphere full of life and deep emotions.", + "seed": 26266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_15/index_3776_seed_26266_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.24115167623385786, + "vlm_logprob": -0.24115167623385786, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.24115167623385786, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0286042969673872 + ], + "opposite_logprob_tokens": [ + -3.77860426902771 + ], + "logprob_reward": -0.0286042969673872, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0286042969673872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6693859696388245, + 0.020950865000486374, + 0.06381548941135406, + 0.01685129851102829, + 0.014771994203329086, + 0.0350988432765007, + 0.030403651297092438, + 0.11992238461971283, + 0.02347514033317566, + 0.005324399098753929 + ], + "credit_map": [ + 0.06631161272525787, + 0.10578842461109161, + 0.11946087330579758, + 0.13678361475467682, + 0.0868673250079155, + 0.03849491477012634, + 0.07894205302000046, + 0.036737110465765, + 0.0018205802189186215, + 0.0022370684891939163, + 0.0026602898724377155, + 0.003327886573970318, + 0.002234372776001692, + 0.004286342766135931, + 0.0026528569869697094, + 0.001731467549689114, + 0.006351675372570753, + 0.012713783420622349, + 0.008983827196061611, + 0.006894670892506838, + 0.007889792323112488, + 0.009117526933550835, + 0.00691085122525692, + 0.004953359719365835, + 0.0014599973801523447, + 0.00148392200935632, + 0.0041650207713246346, + 0.001466180314309895, + 0.0037158627528697252, + 0.0011463561095297337, + 0.002036292338743806, + 0.0013776675332337618, + 0.0012621597852557898, + 0.001931202132254839, + 0.0020400492940098047, + 0.003108226228505373, + 0.0018025292083621025, + 0.001304451609030366, + 0.001955547137185931, + 0.0013678286923095584, + 0.0051707923412323, + 0.001834623864851892, + 0.0017484716372564435, + 0.0018821967532858253, + 0.012134075164794922, + 0.0030374464113265276, + 0.00259657041169703, + 0.006694667506963015, + 0.0023150350898504257, + 0.003186644986271858, + 0.0041333832778036594, + 0.002812193939462304, + 0.0058850753121078014, + 0.0021946525666862726, + 0.004536268766969442, + 0.0053403982892632484, + 0.01434844359755516, + 0.005196854937821627, + 0.03608237951993942, + 0.006568741984665394, + 0.03161795437335968, + 0.0054620704613626, + 0.007190193049609661, + 0.013455749489367008, + 0.0016877420712262392, + 0.0015345355495810509, + 0.004108321387320757, + 0.0012636365136131644, + 0.006008841563016176, + 0.0019050015835091472, + 0.0027954704128205776, + 0.004171591717749834, + 0.0008467223378829658, + 0.00047106394777074456, + 0.0006449049105867743, + 0.000621045648586005, + 0.0007742520538158715, + 0.0006700681406073272, + 0.0006465094047598541, + 0.0006498328293673694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0286042969673872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539902448654175 + ], + "opposite_logprob_tokens": [ + -0.4289902448654175 + ], + "logprob_reward": -1.0539902448654175, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539902448654175, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.37687602639198303, + 0.089000403881073, + 0.10556396096944809, + 0.05700254440307617, + 0.02782256342470646, + 0.056093212217092514, + 0.03643571957945824, + 0.1923486292362213, + 0.03984197974205017, + 0.019014997407794 + ], + "credit_map": [ + 0.04001807048916817, + 0.059465743601322174, + 0.06187284737825394, + 0.07803041487932205, + 0.05000590533018112, + 0.02433454804122448, + 0.04156339913606644, + 0.021585116162896156, + 0.008414042182266712, + 0.005924778524786234, + 0.026537148281931877, + 0.009760378859937191, + 0.015153131447732449, + 0.0075295958667993546, + 0.011625491082668304, + 0.004055836703628302, + 0.00978759489953518, + 0.02464868500828743, + 0.01150648295879364, + 0.009738043881952763, + 0.01803923025727272, + 0.0141405975446105, + 0.010341411456465721, + 0.007361914962530136, + 0.007749801501631737, + 0.002786308294162154, + 0.019261086359620094, + 0.004374886862933636, + 0.013069996610283852, + 0.0032025736290961504, + 0.0025111609138548374, + 0.004046731628477573, + 0.002668422181159258, + 0.003109120763838291, + 0.0043952567502856255, + 0.004709908273071051, + 0.00342861982062459, + 0.0025075897574424744, + 0.003542390652000904, + 0.003461257554590702, + 0.008320895954966545, + 0.002925439737737179, + 0.0027908645570278168, + 0.0018927091732621193, + 0.018956826999783516, + 0.004898251499980688, + 0.005264845211058855, + 0.011043377220630646, + 0.0035000231582671404, + 0.0036214764695614576, + 0.004296071361750364, + 0.002504792297258973, + 0.008336649276316166, + 0.002998277312144637, + 0.004282477777451277, + 0.006895950064063072, + 0.018445007503032684, + 0.007971996441483498, + 0.04669882729649544, + 0.010149975307285786, + 0.058251313865184784, + 0.009485929273068905, + 0.012519308365881443, + 0.028826262801885605, + 0.005227699875831604, + 0.0019025986548513174, + 0.009498614817857742, + 0.002148339757695794, + 0.007251436822116375, + 0.0023597474209964275, + 0.004722522106021643, + 0.006731019821017981, + 0.0020165673922747374, + 0.0015648531261831522, + 0.002418369986116886, + 0.002299004467204213, + 0.002416415838524699, + 0.0024730113800615072, + 0.0036430975887924433, + 0.002183678327128291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -1.0539902448654175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747792940586805 + ], + "opposite_logprob_tokens": [ + -5.006747722625732 + ], + "logprob_reward": -0.006747792940586805, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747792940586805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5334802269935608, + 0.044500771909952164, + 0.029169803485274315, + 0.05156460404396057, + 0.01985962875187397, + 0.08743339776992798, + 0.08355683833360672, + 0.06752951443195343, + 0.07038003951311111, + 0.012525211088359356 + ], + "credit_map": [ + 0.053164709359407425, + 0.08534247428178787, + 0.08954121917486191, + 0.11121536046266556, + 0.0665198415517807, + 0.03134991228580475, + 0.06576032936573029, + 0.03058636002242565, + 0.0039093890227377415, + 0.0030924228485673666, + 0.012065090239048004, + 0.004738036543130875, + 0.007653567939996719, + 0.005011711735278368, + 0.005597920157015324, + 0.0024326310958713293, + 0.0030358151998370886, + 0.0037404869217425585, + 0.005779888015240431, + 0.003346268553286791, + 0.00483769690617919, + 0.003122657537460327, + 0.0034443316981196404, + 0.0018626569071784616, + 0.0046390146017074585, + 0.0041603692807257175, + 0.01237548515200615, + 0.0037496304139494896, + 0.01013929769396782, + 0.0038492430467158556, + 0.007293465081602335, + 0.005358091089874506, + 0.0020197241101413965, + 0.002264959504827857, + 0.002759075490757823, + 0.003564173588529229, + 0.0031395326368510723, + 0.0018443296430632472, + 0.002172081731259823, + 0.0020957519300282, + 0.009891994297504425, + 0.0047920141369104385, + 0.006144230719655752, + 0.003759747603908181, + 0.027752455323934555, + 0.007629564497619867, + 0.00883457064628601, + 0.018628815189003944, + 0.005438952706754208, + 0.004272254649549723, + 0.025356657803058624, + 0.010246388614177704, + 0.013747235760092735, + 0.0033766503911465406, + 0.009522649459540844, + 0.011596047319471836, + 0.0053961919620633125, + 0.003295526374131441, + 0.007969316095113754, + 0.0045378319919109344, + 0.020181093364953995, + 0.0056516677141189575, + 0.007938994094729424, + 0.012558885850012302, + 0.009070486761629581, + 0.0031514023430645466, + 0.01941191591322422, + 0.00465205917134881, + 0.018184667453169823, + 0.004911849740892649, + 0.004077283665537834, + 0.006920363754034042, + 0.0014012716710567474, + 0.0014605813194066286, + 0.0013166957069188356, + 0.0013582053361460567, + 0.0018601387273520231, + 0.0013895300216972828, + 0.001878008944913745, + 0.0018607790116220713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.006747792940586805 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033911466598511 + ], + "opposite_logprob_tokens": [ + -2.35033917427063 + ], + "logprob_reward": -0.10033911466598511, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033911466598511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4793442487716675, + 0.02384824864566326, + 0.043583326041698456, + 0.02839989773929119, + 0.012979288585484028, + 0.10988776385784149, + 0.04190477728843689, + 0.2213866412639618, + 0.02643175981938839, + 0.01223404798656702 + ], + "credit_map": [ + 0.046746060252189636, + 0.07580974698066711, + 0.08945418149232864, + 0.09611692279577255, + 0.0624026320874691, + 0.0268420223146677, + 0.05520985648036003, + 0.02676285244524479, + 0.00213841674849391, + 0.002000359818339348, + 0.00453407084569335, + 0.0024854065850377083, + 0.0038634552620351315, + 0.004374005366116762, + 0.0027081097941845655, + 0.0017444262048229575, + 0.003422724548727274, + 0.006273687817156315, + 0.004692140966653824, + 0.0036161139141768217, + 0.007687107659876347, + 0.008785118348896503, + 0.005223426967859268, + 0.0038830109406262636, + 0.002295777201652527, + 0.002413882641121745, + 0.005456212908029556, + 0.002914073644205928, + 0.005731723736971617, + 0.0026711495593190193, + 0.0038818607572466135, + 0.003035219619050622, + 0.001034882152453065, + 0.0013316593831405044, + 0.001702029607258737, + 0.0019469241378828883, + 0.0015981735195964575, + 0.0017015456687659025, + 0.0019509817939251661, + 0.0017130938358604908, + 0.016028020530939102, + 0.005921059288084507, + 0.005670199170708656, + 0.0033004703000187874, + 0.03946083039045334, + 0.008866283111274242, + 0.008719339966773987, + 0.02192157506942749, + 0.0027618147432804108, + 0.005297086201608181, + 0.006805747747421265, + 0.004812348634004593, + 0.006417032331228256, + 0.0028689512982964516, + 0.006439561489969492, + 0.006502238102257252, + 0.024685949087142944, + 0.01136756967753172, + 0.06080450490117073, + 0.011077530682086945, + 0.05922059342265129, + 0.01068638451397419, + 0.015486393123865128, + 0.028057744726538658, + 0.004978461191058159, + 0.001214961288496852, + 0.008670470677316189, + 0.0018514919793233275, + 0.0033499288838356733, + 0.0013399544404819608, + 0.0026499275118112564, + 0.002376565942540765, + 0.0015969807282090187, + 0.0009559966274537146, + 0.0016716915415599942, + 0.001646782853640616, + 0.0015105728525668383, + 0.000985761289484799, + 0.001981748966500163, + 0.0018845141166821122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.10033911466598511 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076931729912758 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.016076931729912758, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076931729912758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07678671181201935, + 0.0688830241560936, + 0.1615990549325943, + 0.029343774542212486, + 0.0315052792429924, + 0.034063052386045456, + 0.07288695871829987, + 0.30879056453704834, + 0.0981782078742981, + 0.11796344816684723 + ], + "credit_map": [ + 0.009797868318855762, + 0.008653565309941769, + 0.011076392605900764, + 0.010593951679766178, + 0.00932552944868803, + 0.006623269524425268, + 0.009241551160812378, + 0.011474572122097015, + 0.006478339433670044, + 0.005235630087554455, + 0.019396759569644928, + 0.007437346037477255, + 0.01159738376736641, + 0.008450957015156746, + 0.0064979493618011475, + 0.0037886479403823614, + 0.015367632731795311, + 0.03165076673030853, + 0.024912532418966293, + 0.015267547219991684, + 0.02206617221236229, + 0.023865176364779472, + 0.015866264700889587, + 0.012602929025888443, + 0.003431811463087797, + 0.0036257305182516575, + 0.00373460934497416, + 0.00274857971817255, + 0.0040436615236103535, + 0.0029184515587985516, + 0.004934751428663731, + 0.003906174562871456, + 0.0027925530448555946, + 0.0038590990006923676, + 0.004741250071674585, + 0.006131342146545649, + 0.003347822930663824, + 0.002929345704615116, + 0.004680315498262644, + 0.0030235436279326677, + 0.002119690179824829, + 0.0016883432399481535, + 0.0055899606086313725, + 0.004187717568129301, + 0.007251649163663387, + 0.0028041505720466375, + 0.004007996525615454, + 0.0064135389402508736, + 0.007034666836261749, + 0.0076292441226542, + 0.01392108853906393, + 0.00890440121293068, + 0.01376765314489603, + 0.004578282590955496, + 0.007417631335556507, + 0.009633978828787804, + 0.03125793859362602, + 0.013840388506650925, + 0.0813671424984932, + 0.017299555242061615, + 0.08933719247579575, + 0.016745995730161667, + 0.016404539346694946, + 0.04253778234124184, + 0.011091535910964012, + 0.004534419160336256, + 0.01166491862386465, + 0.004040208179503679, + 0.032295506447553635, + 0.007907183840870857, + 0.00907463300973177, + 0.017569785937666893, + 0.009034669026732445, + 0.0100125502794981, + 0.008171860128641129, + 0.014416655525565147, + 0.020963216200470924, + 0.015421888791024685, + 0.021064508706331253, + 0.01887807995080948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.016076931729912758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly realistic slow-motion close-up, the camera held still, captures from eye level the moment when billowing white steam slowly rises from a traditional bamboo steamer. This warm, moist vapor gently expands and swirls in the air, delicately intertwining with the cool morning mist seeping in through the window, forming a hazy, dynamic curtain. Within this shrouded haze, the gentle profile of a middle-aged mother emerges, her raven hair pulled back in a simple bun. A few sweat-soaked strands cling to her cheeks, and a few grains of glistening rice, unnoticed by her busy schedule, are clearly visible at her temples. The soft morning light penetrates the mist, outlining a soft edge of light for her figure and making the few grains of rice slightly reflect light. Her focused and serene expression appears particularly moving in the steam that is sometimes thick and sometimes thin. The whole picture, with its blend of warm and cool tones, creates an extremely tranquil and nostalgic atmosphere full of life and deep emotions.", + "seed": 26267, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_15/index_3776_seed_26267_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.26309789139777423, + "vlm_logprob": -0.26309789139777423, + "vlm_yes_no_margin": 3.024999964237213, + "vlm_token_credit": -0.26309789139777423, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023803066462278366 + ], + "opposite_logprob_tokens": [ + -3.8988029956817627 + ], + "logprob_reward": -0.023803066462278366, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023803066462278366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2270355522632599, + 0.06175174191594124, + 0.021860670298337936, + 0.24795691668987274, + 0.06225767731666565, + 0.0658198893070221, + 0.26647570729255676, + 0.012557276524603367, + 0.01607997715473175, + 0.018204569816589355 + ], + "credit_map": [ + 0.01808592490851879, + 0.03279927372932434, + 0.033686548471450806, + 0.04721962660551071, + 0.029863623902201653, + 0.022486761212348938, + 0.02586614154279232, + 0.017027655616402626, + 0.00577722629532218, + 0.009683862328529358, + 0.00620667077600956, + 0.011793953366577625, + 0.00852228980511427, + 0.005847786087542772, + 0.006800401024520397, + 0.007119548041373491, + 0.00236546085216105, + 0.0031075605656951666, + 0.0025111795403063297, + 0.002929137786850333, + 0.003676564898341894, + 0.002681860700249672, + 0.0025241237599402666, + 0.0020647821947932243, + 0.018924299627542496, + 0.09307463467121124, + 0.07249435037374496, + 0.01172531209886074, + 0.027804149314761162, + 0.0039610364474356174, + 0.014767790213227272, + 0.005205338820815086, + 0.004341733176261187, + 0.015427973121404648, + 0.004413002636283636, + 0.007351824082434177, + 0.009566949680447578, + 0.0052772993221879005, + 0.0100599629804492, + 0.005818933714181185, + 0.014622711576521397, + 0.010857663117349148, + 0.014944708906114101, + 0.004984493367373943, + 0.004974915646016598, + 0.002832663943991065, + 0.008159455843269825, + 0.004443272482603788, + 0.020300183445215225, + 0.0752299353480339, + 0.01852412335574627, + 0.028148451820015907, + 0.04883747175335884, + 0.017516784369945526, + 0.037871766835451126, + 0.02004697360098362, + 0.0012640280183404684, + 0.002334544202312827, + 0.0013201660476624966, + 0.0021136209834367037, + 0.0019106160616502166, + 0.0008994342060759664, + 0.0019413760164752603, + 0.0007734915125183761, + 0.0008749788394197822, + 0.00423781480640173, + 0.0007697994005866349, + 0.0021214857697486877, + 0.003151849377900362, + 0.0009648489649407566, + 0.002610519528388977, + 0.001348680816590786, + 0.0011482120025902987, + 0.0032124598510563374, + 0.0016285576857626438, + 0.0026660733856260777, + 0.0029716521967202425, + 0.001170613570138812, + 0.0037038063164800406, + 0.0017031937604770064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.023803066462278366 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236676216125488 + ], + "opposite_logprob_tokens": [ + -0.3486676812171936 + ], + "logprob_reward": -1.2236676216125488, + "yes_no_margin_reward": -0.8749999403953552, + "token_credit_reward": -1.2236676216125488, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03653690218925476, + 0.0693240612745285, + 0.0485190823674202, + 0.06892582029104233, + 0.02061786875128746, + 0.04310264810919762, + 0.401806116104126, + 0.11454066634178162, + 0.09011458605527878, + 0.10651227086782455 + ], + "credit_map": [ + 0.0037165703251957893, + 0.00491824047639966, + 0.005051237531006336, + 0.005680631380528212, + 0.003217472694814205, + 0.005157805979251862, + 0.0037939948961138725, + 0.005000943783670664, + 0.006900923326611519, + 0.014047064818441868, + 0.011448346078395844, + 0.011655651964247227, + 0.009407740086317062, + 0.004674366209656, + 0.005937142297625542, + 0.005252820439636707, + 0.004263358656316996, + 0.011490652337670326, + 0.004596508573740721, + 0.005353433545678854, + 0.005594007205218077, + 0.0041034952737390995, + 0.005694667808711529, + 0.007422956638038158, + 0.007729720324277878, + 0.011499404907226562, + 0.008256785571575165, + 0.009605462662875652, + 0.009892737492918968, + 0.006466588471084833, + 0.009131895378232002, + 0.006343222688883543, + 0.0015096331480890512, + 0.0034999200142920017, + 0.0017999589908868074, + 0.0032642099540680647, + 0.0028656001668423414, + 0.0014738584868609905, + 0.0034754439257085323, + 0.0027292422018945217, + 0.007133904378861189, + 0.00516414362937212, + 0.006948667112737894, + 0.005843354854732752, + 0.00539233535528183, + 0.0027098467107862234, + 0.005647522863000631, + 0.004262869246304035, + 0.03243453428149223, + 0.10936674475669861, + 0.028510624542832375, + 0.04636077210307121, + 0.07320818305015564, + 0.027526551857590675, + 0.05508732050657272, + 0.02931136079132557, + 0.005762381944805384, + 0.021041415631771088, + 0.006998451892286539, + 0.026832416653633118, + 0.016350341960787773, + 0.005548972170799971, + 0.02476458065211773, + 0.007242097519338131, + 0.004093227442353964, + 0.016635026782751083, + 0.004413678776472807, + 0.01860680989921093, + 0.017217060551047325, + 0.004100052639842033, + 0.01703806035220623, + 0.008010664954781532, + 0.0060257138684391975, + 0.015039189718663692, + 0.00701510114595294, + 0.017312314361333847, + 0.016008051112294197, + 0.006218252703547478, + 0.0175321027636528, + 0.021361542865633965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -1.2236676216125488 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742938421666622 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006742938421666622, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742938421666622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12679094076156616, + 0.07966993749141693, + 0.1225312352180481, + 0.10795584321022034, + 0.058142341673374176, + 0.14294888079166412, + 0.12472912669181824, + 0.11485998332500458, + 0.03231356665492058, + 0.09005817025899887 + ], + "credit_map": [ + 0.010148590430617332, + 0.017989717423915863, + 0.016795674338936806, + 0.029586488381028175, + 0.015304360538721085, + 0.012512749992311, + 0.015596822835505009, + 0.008856532163918018, + 0.007767824921756983, + 0.012307264842092991, + 0.01012000348418951, + 0.012133132666349411, + 0.010918899439275265, + 0.008671259507536888, + 0.009643886238336563, + 0.00810766126960516, + 0.011300223879516125, + 0.03677784278988838, + 0.02156497910618782, + 0.012173487804830074, + 0.013985322788357735, + 0.005289201159030199, + 0.014905722811818123, + 0.0065344590693712234, + 0.011502248235046864, + 0.02117353491485119, + 0.014035922475159168, + 0.013055043295025826, + 0.014889627695083618, + 0.007879328913986683, + 0.012566777877509594, + 0.012853345833718777, + 0.006431511603295803, + 0.009699800051748753, + 0.004878052044659853, + 0.00985577143728733, + 0.00925034936517477, + 0.0034262710250914097, + 0.00949475634843111, + 0.0051058270037174225, + 0.009388257749378681, + 0.04792434349656105, + 0.005244000349193811, + 0.012291437946259975, + 0.017228052020072937, + 0.0059618232771754265, + 0.031225748360157013, + 0.013685212470591068, + 0.005654505919665098, + 0.03562512248754501, + 0.005356507375836372, + 0.017268555238842964, + 0.03259725123643875, + 0.005714161321520805, + 0.013209840282797813, + 0.009303181432187557, + 0.005031532607972622, + 0.019800720736384392, + 0.006040151230990887, + 0.027420857921242714, + 0.014432297088205814, + 0.009242395870387554, + 0.02511671371757984, + 0.0077753267250955105, + 0.001998636871576309, + 0.005555564537644386, + 0.002048071939498186, + 0.007424243725836277, + 0.00536733353510499, + 0.0018772456096485257, + 0.005471584852784872, + 0.0025708850007504225, + 0.004809780046343803, + 0.013701913878321648, + 0.006121065001934767, + 0.015821143984794617, + 0.013356775045394897, + 0.005660077091306448, + 0.015114468522369862, + 0.015472948551177979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.006742938421666622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307904466986656 + ], + "opposite_logprob_tokens": [ + -3.168079137802124 + ], + "logprob_reward": -0.04307904466986656, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307904466986656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16456586122512817, + 0.04939129948616028, + 0.10772872716188431, + 0.10931701213121414, + 0.06377429515123367, + 0.09259352833032608, + 0.22849038243293762, + 0.0794903114438057, + 0.026445206254720688, + 0.07820336520671844 + ], + "credit_map": [ + 0.017149001359939575, + 0.021279143169522285, + 0.02521313726902008, + 0.03001488372683525, + 0.020463259890675545, + 0.015292661264538765, + 0.01913098432123661, + 0.016022777184844017, + 0.0050519416108727455, + 0.006767477374523878, + 0.005951176397502422, + 0.008257229812443256, + 0.006253449246287346, + 0.005520912818610668, + 0.004943063482642174, + 0.0066460492089390755, + 0.015018393285572529, + 0.021725492551922798, + 0.017965706065297127, + 0.014681695029139519, + 0.014173702336847782, + 0.007612124085426331, + 0.01056923158466816, + 0.005982392933219671, + 0.008281699381768703, + 0.03166574612259865, + 0.020503632724285126, + 0.010488977655768394, + 0.014185546897351742, + 0.0053369393572211266, + 0.010071998462080956, + 0.008782478980720043, + 0.00434328056871891, + 0.010309365577995777, + 0.006774153560400009, + 0.00790758989751339, + 0.014260630123317242, + 0.005187700968235731, + 0.00893842801451683, + 0.006053154822438955, + 0.024049270898103714, + 0.013495489954948425, + 0.02005406655371189, + 0.008277223445475101, + 0.007182118017226458, + 0.0031227024737745523, + 0.009818815626204014, + 0.006593849044293165, + 0.015842339023947716, + 0.06344935297966003, + 0.01638197898864746, + 0.029405834153294563, + 0.04077272489666939, + 0.018985534086823463, + 0.029142191633582115, + 0.014510422945022583, + 0.003726138500496745, + 0.01359342597424984, + 0.004188690334558487, + 0.01771780662238598, + 0.01074313186109066, + 0.0062593454495072365, + 0.01719435304403305, + 0.006067420821636915, + 0.0014485750580206513, + 0.004961415193974972, + 0.0015606684610247612, + 0.004867230076342821, + 0.005257181357592344, + 0.0015431049978360534, + 0.00460920250043273, + 0.0021978297736495733, + 0.0037859089206904173, + 0.01264380943030119, + 0.005147016607224941, + 0.01808919571340084, + 0.012696536257863045, + 0.005115260370075703, + 0.013374482281506062, + 0.0073511600494384766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.04307904466986656 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819678582251072 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.01819678582251072, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819678582251072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10662281513214111, + 0.12989595532417297, + 0.060944292694330215, + 0.14118944108486176, + 0.07922068983316422, + 0.06760524958372116, + 0.11370962113142014, + 0.09606251120567322, + 0.08804579824209213, + 0.11670368909835815 + ], + "credit_map": [ + 0.008613822981715202, + 0.012787917628884315, + 0.013219726271927357, + 0.016625259071588516, + 0.018803009763360023, + 0.010435882955789566, + 0.014684576541185379, + 0.011452616192400455, + 0.0120448162779212, + 0.025057118386030197, + 0.02408020943403244, + 0.020536210387945175, + 0.016487967222929, + 0.008217750117182732, + 0.015029351226985455, + 0.008442528545856476, + 0.006585892755538225, + 0.01323522999882698, + 0.007343172561377287, + 0.008381692692637444, + 0.009085966274142265, + 0.004959348123520613, + 0.00714956084266305, + 0.004203427582979202, + 0.009466488845646381, + 0.04592307657003403, + 0.034429021179676056, + 0.012150613591074944, + 0.016846049576997757, + 0.005011696368455887, + 0.011700763367116451, + 0.005661729257553816, + 0.006429500877857208, + 0.012759104371070862, + 0.009769145399332047, + 0.015212242491543293, + 0.009780517779290676, + 0.006867205258458853, + 0.012602822855114937, + 0.005800147540867329, + 0.01624281331896782, + 0.007001429330557585, + 0.013887769542634487, + 0.008285594172775745, + 0.006295819766819477, + 0.0025973014999181032, + 0.007725799921900034, + 0.005568727385252714, + 0.013014898635447025, + 0.018192250281572342, + 0.009905636310577393, + 0.016466261819005013, + 0.017304766923189163, + 0.012301339767873287, + 0.014379474334418774, + 0.012144995853304863, + 0.007223109714686871, + 0.016497410833835602, + 0.00553609523922205, + 0.015860505402088165, + 0.01906980760395527, + 0.011725091375410557, + 0.012489782646298409, + 0.007660709787160158, + 0.004883036948740482, + 0.016092130914330482, + 0.004378631245344877, + 0.017952999100089073, + 0.0178273506462574, + 0.005345073062926531, + 0.013033348135650158, + 0.008533226326107979, + 0.008770778775215149, + 0.015566994436085224, + 0.010791872628033161, + 0.023757293820381165, + 0.016158243641257286, + 0.007356313522905111, + 0.02090892754495144, + 0.013393256813287735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.01819678582251072 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly realistic slow-motion close-up, the camera held still, captures from eye level the moment when billowing white steam slowly rises from a traditional bamboo steamer. This warm, moist vapor gently expands and swirls in the air, delicately intertwining with the cool morning mist seeping in through the window, forming a hazy, dynamic curtain. Within this shrouded haze, the gentle profile of a middle-aged mother emerges, her raven hair pulled back in a simple bun. A few sweat-soaked strands cling to her cheeks, and a few grains of glistening rice, unnoticed by her busy schedule, are clearly visible at her temples. The soft morning light penetrates the mist, outlining a soft edge of light for her figure and making the few grains of rice slightly reflect light. Her focused and serene expression appears particularly moving in the steam that is sometimes thick and sometimes thin. The whole picture, with its blend of warm and cool tones, creates an extremely tranquil and nostalgic atmosphere full of life and deep emotions.", + "seed": 26268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_15/index_3776_seed_26268_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.24797644577920436, + "vlm_logprob": -0.24797644577920436, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.24797644577920436, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014755051583051682 + ], + "opposite_logprob_tokens": [ + -4.3897552490234375 + ], + "logprob_reward": -0.014755051583051682, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014755051583051682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16507261991500854, + 0.047072019428014755, + 0.03255395218729973, + 0.02925020083785057, + 0.054210975766181946, + 0.07629325985908508, + 0.3546602427959442, + 0.1595020294189453, + 0.04053555801510811, + 0.04084914177656174 + ], + "credit_map": [ + 0.019654661417007446, + 0.026768531650304794, + 0.0246669240295887, + 0.02490188553929329, + 0.02641301043331623, + 0.011469239369034767, + 0.017361905425786972, + 0.013836442492902279, + 0.005021079443395138, + 0.0077594853937625885, + 0.007961414754390717, + 0.007012607995420694, + 0.005942182615399361, + 0.004853257443755865, + 0.004744925070554018, + 0.0037770678754895926, + 0.0034553755540400743, + 0.005275797564536333, + 0.002778847236186266, + 0.0036302278749644756, + 0.004210175946354866, + 0.0049109021201729774, + 0.005208055954426527, + 0.0030845701694488525, + 0.0027552794199436903, + 0.006624894216656685, + 0.0044882772490382195, + 0.003069041995331645, + 0.0035313458647578955, + 0.0030206686351448298, + 0.003772065043449402, + 0.0019886286463588476, + 0.0031507567036896944, + 0.012508293613791466, + 0.004243850242346525, + 0.007018627133220434, + 0.010839446447789669, + 0.003744641086086631, + 0.0065471092239022255, + 0.006158254109323025, + 0.0053531150333583355, + 0.009212897159159184, + 0.009582108817994595, + 0.010741605423390865, + 0.014199267141520977, + 0.006330507341772318, + 0.012509887106716633, + 0.008363871835172176, + 0.014375499449670315, + 0.024756068363785744, + 0.08190006762742996, + 0.09448838233947754, + 0.03421395644545555, + 0.02760106883943081, + 0.02739677205681801, + 0.0499284602701664, + 0.006943175103515387, + 0.007836529053747654, + 0.05028131976723671, + 0.027047300711274147, + 0.010258941911160946, + 0.012107455171644688, + 0.00977737084031105, + 0.03524993732571602, + 0.0044703492894768715, + 0.00373479095287621, + 0.006021502893418074, + 0.0032367275562137365, + 0.006582124624401331, + 0.0038714741822332144, + 0.006268055643886328, + 0.0063505335710942745, + 0.0032901777885854244, + 0.004104012157768011, + 0.00560007942840457, + 0.0041399491019546986, + 0.008050774224102497, + 0.004948665853589773, + 0.0036454477813094854, + 0.00707003241404891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.014755051583051682 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.13707435131073 + ], + "opposite_logprob_tokens": [ + -0.38707438111305237 + ], + "logprob_reward": -1.13707435131073, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.13707435131073, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1283261477947235, + 0.09953135997056961, + 0.14733287692070007, + 0.13447552919387817, + 0.08182007819414139, + 0.07353639602661133, + 0.09722818434238434, + 0.1409764140844345, + 0.05930975079536438, + 0.03746325150132179 + ], + "credit_map": [ + 0.009746355935931206, + 0.018040824681520462, + 0.02789577841758728, + 0.02248937450349331, + 0.01637408323585987, + 0.008064432069659233, + 0.013982512056827545, + 0.01173277385532856, + 0.01088500116020441, + 0.019017167389392853, + 0.011392259038984776, + 0.012747889384627342, + 0.015078765340149403, + 0.008304760791361332, + 0.014429875649511814, + 0.007675647269934416, + 0.019189197570085526, + 0.02135179378092289, + 0.011587738059461117, + 0.019008347764611244, + 0.018906306475400925, + 0.013666346669197083, + 0.028509799391031265, + 0.015113357454538345, + 0.01457237359136343, + 0.02323654666543007, + 0.022857969626784325, + 0.013961018063127995, + 0.016769424080848694, + 0.01829739473760128, + 0.014818210154771805, + 0.009962601587176323, + 0.005202769301831722, + 0.020118599757552147, + 0.017169423401355743, + 0.010429639369249344, + 0.009601600468158722, + 0.0051567042246460915, + 0.007194853853434324, + 0.006946480367332697, + 0.00579679012298584, + 0.00749746710062027, + 0.007610362023115158, + 0.008869858458638191, + 0.014882883988320827, + 0.009407871402800083, + 0.009806441143155098, + 0.009664724580943584, + 0.0065498738549649715, + 0.007826101034879684, + 0.0294113177806139, + 0.010055902414023876, + 0.012119017541408539, + 0.007025865837931633, + 0.00961711909621954, + 0.014622990973293781, + 0.00828523002564907, + 0.007498613558709621, + 0.04519720748066902, + 0.025706565007567406, + 0.009930161759257317, + 0.010280015878379345, + 0.010238079354166985, + 0.02384052239358425, + 0.004723805468529463, + 0.006612798199057579, + 0.009472827427089214, + 0.0050773611292243, + 0.007960728369653225, + 0.007128201890736818, + 0.006874085403978825, + 0.01145994197577238, + 0.004322021268308163, + 0.003925069235265255, + 0.006250815466046333, + 0.0036566813942044973, + 0.005877153016626835, + 0.003929335158318281, + 0.00402447022497654, + 0.00547770643606782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -1.13707435131073 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744714453816414 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.006744714453816414, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744714453816414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03487314283847809, + 0.04605301469564438, + 0.05272383615374565, + 0.018731696531176567, + 0.1428055614233017, + 0.1738583743572235, + 0.2582799196243286, + 0.09865844249725342, + 0.13515624403953552, + 0.0388597771525383 + ], + "credit_map": [ + 0.004496229346841574, + 0.005001003388315439, + 0.005462840665131807, + 0.005587482359260321, + 0.004681650083512068, + 0.0016634833300486207, + 0.004623343702405691, + 0.003357108449563384, + 0.0034805068280547857, + 0.007159762550145388, + 0.0067147789523005486, + 0.006025385111570358, + 0.005888199433684349, + 0.0029591848142445087, + 0.009291316382586956, + 0.004533880390226841, + 0.004663769155740738, + 0.00511027779430151, + 0.00307659525424242, + 0.006405828986316919, + 0.00926552526652813, + 0.004749269224703312, + 0.013799098320305347, + 0.005653473548591137, + 0.0018757794750854373, + 0.0037335455417633057, + 0.0017893188633024693, + 0.0019996354822069407, + 0.0018465635366737843, + 0.0022286258172243834, + 0.0024521909654140472, + 0.0028060374315828085, + 0.009140212088823318, + 0.026900433003902435, + 0.010464652441442013, + 0.012759819626808167, + 0.0401478037238121, + 0.011286193504929543, + 0.01061138603836298, + 0.021495074033737183, + 0.011126063764095306, + 0.020563693717122078, + 0.02768081985414028, + 0.0168040469288826, + 0.030383430421352386, + 0.02188447117805481, + 0.02582661807537079, + 0.019589228555560112, + 0.011166608892381191, + 0.019292257726192474, + 0.03072345070540905, + 0.06446447223424911, + 0.031770285218954086, + 0.025493241846561432, + 0.02180042862892151, + 0.05356915667653084, + 0.004304560832679272, + 0.006917459890246391, + 0.02949153259396553, + 0.016607578843832016, + 0.008095032535493374, + 0.0076201967895030975, + 0.0066549829207360744, + 0.018967101350426674, + 0.011163550429046154, + 0.013350357301533222, + 0.013390595093369484, + 0.012454086914658546, + 0.024994544684886932, + 0.013986333273351192, + 0.023216785863041878, + 0.022599978372454643, + 0.003828897373750806, + 0.0031470442190766335, + 0.007801236119121313, + 0.0042907982133328915, + 0.006182867102324963, + 0.0043783471919596195, + 0.004912099801003933, + 0.004318487364798784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.006744714453816414 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495576187968254 + ], + "opposite_logprob_tokens": [ + -2.9299557209014893 + ], + "logprob_reward": -0.05495576187968254, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495576187968254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06452120095491409, + 0.03043324314057827, + 0.061765242367982864, + 0.05566706508398056, + 0.06066862493753433, + 0.11447543650865555, + 0.48755815625190735, + 0.05341189354658127, + 0.040997687727212906, + 0.030501436442136765 + ], + "credit_map": [ + 0.009069832041859627, + 0.01043627131730318, + 0.00960410013794899, + 0.010169988498091698, + 0.009006625972688198, + 0.004117179196327925, + 0.006923995912075043, + 0.005193209741264582, + 0.0040860953740775585, + 0.004239888861775398, + 0.0044378614984452724, + 0.004317421466112137, + 0.0034821387380361557, + 0.0028485124930739403, + 0.004400108940899372, + 0.0026212164666503668, + 0.0072577958926558495, + 0.007873186841607094, + 0.004971911199390888, + 0.009350636973977089, + 0.00718290638178587, + 0.005274413153529167, + 0.013214652426540852, + 0.006639738101512194, + 0.007606207393109798, + 0.012934744358062744, + 0.008702700957655907, + 0.0062605757266283035, + 0.006259859073907137, + 0.005017311777919531, + 0.0057853334583342075, + 0.0031003335025161505, + 0.0031066532246768475, + 0.016527608036994934, + 0.010764451697468758, + 0.008236439898610115, + 0.0074341269209980965, + 0.0037179412320256233, + 0.006451975554227829, + 0.0044294316321611404, + 0.0074804918840527534, + 0.011958948336541653, + 0.014748780056834221, + 0.017791980877518654, + 0.018534379079937935, + 0.014241741970181465, + 0.016864515841007233, + 0.012854594737291336, + 0.019664838910102844, + 0.033459048718214035, + 0.09671517461538315, + 0.1258002072572708, + 0.04683114215731621, + 0.04219622164964676, + 0.0388965904712677, + 0.08399493247270584, + 0.002239413093775511, + 0.0032612879294902086, + 0.016651105135679245, + 0.00935949757695198, + 0.0059675294905900955, + 0.0037541352212429047, + 0.00514888996258378, + 0.007030039560049772, + 0.00346977892331779, + 0.0036175930872559547, + 0.004858509171754122, + 0.0032800938934087753, + 0.008025663904845715, + 0.004096976947039366, + 0.0066733770072460175, + 0.006975692696869373, + 0.0034064834471791983, + 0.0022071937564760447, + 0.00469979690387845, + 0.0027046881150454283, + 0.005669573321938515, + 0.003570981789380312, + 0.0031279714312404394, + 0.005114748142659664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.05495576187968254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a slow-motion close-up of a traditional bamboo steamer emitting billowing white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still at eye level throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged woman with raven hair in a bun visible within the steam haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are glistening grains of rice visible on the woman's temples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature soft morning light creating a rim light effect on the woman's figure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352349668741226 + ], + "opposite_logprob_tokens": [ + -3.6513524055480957 + ], + "logprob_reward": -0.026352349668741226, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352349668741226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0984080359339714, + 0.05029839649796486, + 0.04573317989706993, + 0.0456220842897892, + 0.10490474849939346, + 0.053386710584163666, + 0.3165047764778137, + 0.06628111749887466, + 0.08422711491584778, + 0.13463380932807922 + ], + "credit_map": [ + 0.02240358479321003, + 0.010950500145554543, + 0.013948887586593628, + 0.01585058681666851, + 0.014605289325118065, + 0.005261734593659639, + 0.008150278590619564, + 0.007237170357257128, + 0.0043839155696332455, + 0.007557559758424759, + 0.007379965856671333, + 0.006596817169338465, + 0.007904289290308952, + 0.0038079542573541403, + 0.008213374763727188, + 0.004454517737030983, + 0.004591910634189844, + 0.0069081601686775684, + 0.0037207629065960646, + 0.00537848798558116, + 0.005703459959477186, + 0.006340483669191599, + 0.00793911051005125, + 0.005150801036506891, + 0.004599194973707199, + 0.009250869043171406, + 0.007841940969228745, + 0.0041779098100960255, + 0.005835619289427996, + 0.005325447767972946, + 0.005378213711082935, + 0.0032128903549164534, + 0.00989533681422472, + 0.028812391683459282, + 0.007646162062883377, + 0.013108666986227036, + 0.019371220842003822, + 0.006543076131492853, + 0.007833259180188179, + 0.011694634333252907, + 0.0051671299152076244, + 0.004940647631883621, + 0.006300332956016064, + 0.007893466390669346, + 0.007337022572755814, + 0.005108731798827648, + 0.010544650256633759, + 0.006094729527831078, + 0.012324602343142033, + 0.019097426906228065, + 0.0770181342959404, + 0.07473362982273102, + 0.03367103636264801, + 0.025042714551091194, + 0.026304274797439575, + 0.0483129620552063, + 0.004582473076879978, + 0.008101695217192173, + 0.003943548537790775, + 0.005470664240419865, + 0.013297537341713905, + 0.008577484637498856, + 0.010994603857398033, + 0.011313115246593952, + 0.009076051414012909, + 0.0072697242721915245, + 0.009075620211660862, + 0.0075819166377186775, + 0.012182212434709072, + 0.008128420449793339, + 0.007034630980342627, + 0.023878542706370354, + 0.017128128558397293, + 0.012121417559683323, + 0.028219003230333328, + 0.011467568576335907, + 0.015496672131121159, + 0.011870655231177807, + 0.02415141463279724, + 0.014178944751620293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.026352349668741226 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_2/index_tensor([3081, 3081])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_2/index_tensor([3081, 3081])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1290894f1bb4072bf5006a0a68abcdd4367fcf33 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_2/index_tensor([3081, 3081])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "strikingly dynamic, low-angle, side-on tracking shot follows an elegant black Maybach as it speeds through a rain-soaked city street at night. The camera's focus is pinpointed on the vehicle's front wheels, their high-gloss multi-spoke alloy wheels spinning at high speed in a blur of dynamic light and shadow. The iconic logo embedded in their centers occasionally gleams coldly under the passing streetlights. The wet black tires slice through the puddle at breakneck speed, violently flinging a vast, spectacular fan-shaped curtain of water to the rear and sides. Countless shimmering droplets, momentarily illuminated by the headlights and the shimmering neon lights in the background, glitter like diamonds tossed into the air. In the background, the wet asphalt reflects the cold white light of the streetlights and the iridescent billboards, creating distorted, flowing reflections. The image, dominated by deep black and wet asphalt gray, is punctuated by neon magenta and electric blue, creating a cool, urban atmosphere imbued with power, speed, and luxury.", + "seed": 6313, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_2/index_3081_seed_6313_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04278268963098526, + "vlm_logprob": -0.04278268963098526, + "vlm_yes_no_margin": 3.7249999761581423, + "vlm_token_credit": -0.04278268963098526, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15193304419517517 + ], + "opposite_logprob_tokens": [ + -2.026932954788208 + ], + "logprob_reward": -0.15193304419517517, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.15193304419517517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23676981031894684, + 0.16088314354419708, + 0.09592136740684509, + 0.0583021380007267, + 0.07591000944375992, + 0.032023172825574875, + 0.032060008496046066, + 0.11662019789218903, + 0.12785755097866058, + 0.06365267187356949 + ], + "credit_map": [ + 0.019007699564099312, + 0.039457447826862335, + 0.02509775012731552, + 0.026529699563980103, + 0.025983011350035667, + 0.03361449018120766, + 0.029209129512310028, + 0.037870585918426514, + 0.007262186147272587, + 0.00967546459287405, + 0.007090271916240454, + 0.06048736721277237, + 0.0077713243663311005, + 0.021628163754940033, + 0.03766724839806557, + 0.009301116690039635, + 0.01005934551358223, + 0.010999893769621849, + 0.01187060121446848, + 0.008024347946047783, + 0.014383409172296524, + 0.01328194048255682, + 0.011532478034496307, + 0.01576935686171055, + 0.005222477950155735, + 0.011062195524573326, + 0.006626212038099766, + 0.008303091861307621, + 0.009065682999789715, + 0.006266522686928511, + 0.005174337420612574, + 0.006581616122275591, + 0.011941498145461082, + 0.011261481791734695, + 0.006010770332068205, + 0.009335217997431755, + 0.011674071662127972, + 0.011425605975091457, + 0.00805090181529522, + 0.0062104640528559685, + 0.004614838864654303, + 0.004550272598862648, + 0.0028348653577268124, + 0.003181508742272854, + 0.003463046858087182, + 0.004324955865740776, + 0.005212837364524603, + 0.0038408467080444098, + 0.005249727517366409, + 0.004252842627465725, + 0.0030381560791283846, + 0.003536828560754657, + 0.005515067372471094, + 0.003290363121777773, + 0.0029134955257177353, + 0.004263527225703001, + 0.01886114664375782, + 0.012077202089130878, + 0.017586510628461838, + 0.008395984768867493, + 0.026209112256765366, + 0.01234600692987442, + 0.009438708424568176, + 0.011705522425472736, + 0.022717129439115524, + 0.013142358511686325, + 0.022423148155212402, + 0.012872496619820595, + 0.01812967285513878, + 0.010823281481862068, + 0.01783076673746109, + 0.009918687865138054, + 0.017806584015488625, + 0.006571443751454353, + 0.005179169587790966, + 0.004423487465828657, + 0.012344191782176495, + 0.004705758765339851, + 0.008299942128360271, + 0.004322091583162546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.15193304419517517 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610848680138588 + ], + "opposite_logprob_tokens": [ + -4.141108512878418 + ], + "logprob_reward": -0.01610848680138588, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610848680138588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054563008248806, + 0.09707049280405045, + 0.07177844643592834, + 0.2554059326648712, + 0.06766598671674728, + 0.05945641174912453, + 0.1848122924566269, + 0.04871674254536629, + 0.04781915992498398, + 0.11271148175001144 + ], + "credit_map": [ + 0.006775729823857546, + 0.007074099034070969, + 0.004801054019480944, + 0.006184861529618502, + 0.006441955454647541, + 0.007546695414930582, + 0.009543821215629578, + 0.006194797344505787, + 0.003623177530243993, + 0.005744930822402239, + 0.00457015773281455, + 0.034874752163887024, + 0.004754875786602497, + 0.010769974440336227, + 0.025554200634360313, + 0.007178435102105141, + 0.0056452336721122265, + 0.008561704307794571, + 0.008986408822238445, + 0.0074930679984390736, + 0.008696360513567924, + 0.00859406404197216, + 0.009648766368627548, + 0.014152848161756992, + 0.02746596559882164, + 0.030022546648979187, + 0.024646559730172157, + 0.027408761903643608, + 0.03335576876997948, + 0.01573432795703411, + 0.03740718960762024, + 0.059364814311265945, + 0.010289966128766537, + 0.009532839991152287, + 0.004466741345822811, + 0.00961191114038229, + 0.011720718815922737, + 0.009176230989396572, + 0.006905981805175543, + 0.005961600691080093, + 0.006798419635742903, + 0.00825346540659666, + 0.009664617478847504, + 0.00639787781983614, + 0.005519900005310774, + 0.007667830679565668, + 0.0050689526833593845, + 0.010085354559123516, + 0.035021618008613586, + 0.014981883578002453, + 0.035496119409799576, + 0.015435339882969856, + 0.025304999202489853, + 0.01908963918685913, + 0.01425107941031456, + 0.025231637060642242, + 0.007833830080926418, + 0.006374747026711702, + 0.006574055179953575, + 0.004634136334061623, + 0.00634808000177145, + 0.005905193276703358, + 0.004617038182914257, + 0.006429667584598064, + 0.006636465899646282, + 0.0052202558144927025, + 0.007437702734023333, + 0.004817600827664137, + 0.007366967387497425, + 0.005502874031662941, + 0.004116994794458151, + 0.006720306817442179, + 0.019703175872564316, + 0.011017736978828907, + 0.01211179792881012, + 0.01063704863190651, + 0.01605392061173916, + 0.014587719924747944, + 0.015426325611770153, + 0.01317377295345068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.01610848680138588 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087356135249138 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087356135249138, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087356135249138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18345704674720764, + 0.22994352877140045, + 0.08107024431228638, + 0.04005425050854683, + 0.047730427235364914, + 0.10237139463424683, + 0.10792455077171326, + 0.048241838812828064, + 0.05972608923912048, + 0.09948057681322098 + ], + "credit_map": [ + 0.01484369020909071, + 0.021570201963186264, + 0.01403353177011013, + 0.016063202172517776, + 0.03737277165055275, + 0.026121756061911583, + 0.02071586810052395, + 0.03273603692650795, + 0.005402224138379097, + 0.011854088865220547, + 0.0060964846052229404, + 0.09312216937541962, + 0.008483066223561764, + 0.02907775342464447, + 0.06261155754327774, + 0.013296209275722504, + 0.00804727803915739, + 0.01129968836903572, + 0.007854729890823364, + 0.0068464274518191814, + 0.011779412627220154, + 0.010573087260127068, + 0.011842584237456322, + 0.012827045284211636, + 0.003837962867692113, + 0.00528915598988533, + 0.004051784984767437, + 0.004043608903884888, + 0.005587071180343628, + 0.002907422138378024, + 0.006720923352986574, + 0.00761632202193141, + 0.006944907829165459, + 0.006464100908488035, + 0.004744245205074549, + 0.008507823571562767, + 0.006963915657252073, + 0.005346256773918867, + 0.004713228903710842, + 0.0040459539741277695, + 0.013306145556271076, + 0.011659652926027775, + 0.010963954962790012, + 0.009187059476971626, + 0.014712934382259846, + 0.021653447300195694, + 0.007787509821355343, + 0.013100698590278625, + 0.019197994843125343, + 0.009068218059837818, + 0.019767913967370987, + 0.008823289535939693, + 0.015517782419919968, + 0.012067228555679321, + 0.008400406688451767, + 0.015081717632710934, + 0.00692754378542304, + 0.004394882824271917, + 0.007492814213037491, + 0.0037966228555887938, + 0.00716385617852211, + 0.006851474288851023, + 0.005007406231015921, + 0.006607247982174158, + 0.008576463907957077, + 0.007956764660775661, + 0.009642863646149635, + 0.009307930245995522, + 0.007987120188772678, + 0.004393496084958315, + 0.006504788529127836, + 0.005356667563319206, + 0.02812720276415348, + 0.009549219161272049, + 0.006325410678982735, + 0.006777394562959671, + 0.02837783843278885, + 0.004607266280800104, + 0.010357619263231754, + 0.0053586335852742195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.011087356135249138 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205172657966614 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014205172657966614, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205172657966614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05063830688595772, + 0.06946338713169098, + 0.11543504893779755, + 0.04379327967762947, + 0.10227078944444656, + 0.09374109655618668, + 0.2813930809497833, + 0.08894030749797821, + 0.06932111829519272, + 0.08500366657972336 + ], + "credit_map": [ + 0.005416836589574814, + 0.00882519781589508, + 0.0032177474349737167, + 0.0042052422650158405, + 0.006661893334239721, + 0.007615451235324144, + 0.007906242273747921, + 0.00678968895226717, + 0.00269536511041224, + 0.0038499499205499887, + 0.0026914498303085566, + 0.024431688711047173, + 0.003904581069946289, + 0.008852122351527214, + 0.015373054891824722, + 0.007665170822292566, + 0.010181733407080173, + 0.014051776379346848, + 0.011518711224198341, + 0.010744914412498474, + 0.01370569970458746, + 0.014993460848927498, + 0.017462721094489098, + 0.022776015102863312, + 0.0051330155692994595, + 0.007767112925648689, + 0.004076314624398947, + 0.0047800540924072266, + 0.006571335718035698, + 0.003967961296439171, + 0.003929343074560165, + 0.007568138185888529, + 0.014340875670313835, + 0.01188779529184103, + 0.007608284242451191, + 0.011284922249615192, + 0.017969174310564995, + 0.018416812643408775, + 0.009238766506314278, + 0.011524148285388947, + 0.01373706292361021, + 0.010728204622864723, + 0.00763690797612071, + 0.007434525061398745, + 0.018559623509645462, + 0.01831095479428768, + 0.00788384024053812, + 0.009449965320527554, + 0.05149521306157112, + 0.021904747933149338, + 0.05432575196027756, + 0.024217892438173294, + 0.03753885254263878, + 0.0304438304156065, + 0.021121932193636894, + 0.04034483805298805, + 0.014529041014611721, + 0.009554296731948853, + 0.01303237397223711, + 0.007469262462109327, + 0.013511420227587223, + 0.01088494248688221, + 0.008657233789563179, + 0.011301730759441853, + 0.011041193269193172, + 0.008675660006701946, + 0.010644136928021908, + 0.010117492638528347, + 0.009236516430974007, + 0.006648698356002569, + 0.007267174776643515, + 0.00569023797288537, + 0.019191935658454895, + 0.010492564179003239, + 0.00874501932412386, + 0.006664854008704424, + 0.010133993811905384, + 0.008470574393868446, + 0.012273654341697693, + 0.009031067602336407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.014205172657966614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579388365149498 + ], + "opposite_logprob_tokens": [ + -3.8955793380737305 + ], + "logprob_reward": -0.020579388365149498, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579388365149498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06187550351023674, + 0.164825901389122, + 0.19325539469718933, + 0.07695228606462479, + 0.05196951702237129, + 0.061335355043411255, + 0.18522101640701294, + 0.0509786419570446, + 0.024019703269004822, + 0.12956669926643372 + ], + "credit_map": [ + 0.007714229170233011, + 0.010270318947732449, + 0.005238756071776152, + 0.0078026545234024525, + 0.009202945977449417, + 0.007484861649572849, + 0.0057836174964904785, + 0.00837811827659607, + 0.004541803617030382, + 0.009181982837617397, + 0.004719683900475502, + 0.06562124192714691, + 0.006408000364899635, + 0.019836097955703735, + 0.04461989551782608, + 0.009897207841277122, + 0.017833130434155464, + 0.024401454254984856, + 0.018299216404557228, + 0.020983785390853882, + 0.023718995973467827, + 0.025969581678509712, + 0.024258360266685486, + 0.037790872156620026, + 0.005401985719799995, + 0.008115319535136223, + 0.012224878184497356, + 0.0070130545645952225, + 0.011123181320726871, + 0.005039847455918789, + 0.012403594329953194, + 0.015630429610610008, + 0.007315879222005606, + 0.008678175508975983, + 0.006092841271311045, + 0.007196047808974981, + 0.00706350477412343, + 0.005066243465989828, + 0.005992338992655277, + 0.004564485512673855, + 0.009051031433045864, + 0.0075608971528708935, + 0.008430108428001404, + 0.0062741744332015514, + 0.004293026868253946, + 0.01098867878317833, + 0.006318079773336649, + 0.008419360034167767, + 0.034399595111608505, + 0.015533225610852242, + 0.03575819730758667, + 0.015656476840376854, + 0.024325819686055183, + 0.020315101370215416, + 0.013892666436731815, + 0.02533993497490883, + 0.008380291052162647, + 0.007694647181779146, + 0.007932700216770172, + 0.005900692194700241, + 0.0072355810552835464, + 0.003989371005445719, + 0.004405050538480282, + 0.0054403068497776985, + 0.003295222530141473, + 0.003052398096770048, + 0.004181946627795696, + 0.0046132635325193405, + 0.002881736494600773, + 0.0017624669708311558, + 0.0025724449660629034, + 0.0016602234682068229, + 0.02017621509730816, + 0.01112351380288601, + 0.01091208215802908, + 0.009609417989850044, + 0.028504636138677597, + 0.01205828320235014, + 0.022468721494078636, + 0.014713825657963753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.020579388365149498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic, low-angle, side-on tracking shot follows an elegant black Maybach as it speeds through a rain-soaked city street at night. The camera's focus is pinpointed on the vehicle's front wheels, their high-gloss multi-spoke alloy wheels spinning at high speed in a blur of dynamic light and shadow. The iconic logo embedded in their centers occasionally gleams coldly under the passing streetlights. The wet black tires slice through the puddle at breakneck speed, violently flinging a vast, spectacular fan-shaped curtain of water to the rear and sides. Countless shimmering droplets, momentarily illuminated by the headlights and the shimmering neon lights in the background, glitter like diamonds tossed into the air. In the background, the wet asphalt reflects the cold white light of the streetlights and the iridescent billboards, creating distorted, flowing reflections. The image, dominated by deep black and wet asphalt gray, is punctuated by neon magenta and electric blue, creating a cool, urban atmosphere imbued with power, speed, and luxury.", + "seed": 6314, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_2/index_3081_seed_6314_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03687861114740372, + "vlm_logprob": -0.03687861114740372, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.03687861114740372, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09975709766149521 + ], + "opposite_logprob_tokens": [ + -2.474757194519043 + ], + "logprob_reward": -0.09975709766149521, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09975709766149521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2004743069410324, + 0.07477686554193497, + 0.10305498540401459, + 0.12027093023061752, + 0.1151139885187149, + 0.05464796721935272, + 0.058677446097135544, + 0.04055071622133255, + 0.14090535044670105, + 0.09152743965387344 + ], + "credit_map": [ + 0.027803799137473106, + 0.025613093748688698, + 0.035793330520391464, + 0.042369402945041656, + 0.022394463419914246, + 0.012514844536781311, + 0.02177930623292923, + 0.012206063605844975, + 0.012915799394249916, + 0.009521646425127983, + 0.012791355140507221, + 0.007202471606433392, + 0.00987959560006857, + 0.007225687615573406, + 0.005664383061230183, + 0.009575928561389446, + 0.012862611562013626, + 0.011722929775714874, + 0.011015481315553188, + 0.009734606370329857, + 0.008238586597144604, + 0.01013109553605318, + 0.027333976700901985, + 0.012015698477625847, + 0.013577599078416824, + 0.009387639351189137, + 0.01640351489186287, + 0.010349621064960957, + 0.014861466363072395, + 0.018261868506669998, + 0.020250171422958374, + 0.017179051414132118, + 0.013490200974047184, + 0.009888111613690853, + 0.013959119096398354, + 0.008526247926056385, + 0.016830023378133774, + 0.023785753175616264, + 0.013959635980427265, + 0.014674892649054527, + 0.004048463422805071, + 0.004551031161099672, + 0.007794301491230726, + 0.004379905294626951, + 0.006125297863036394, + 0.011629180051386356, + 0.008734948001801968, + 0.007384840864688158, + 0.0056569622829556465, + 0.006407192908227444, + 0.0057409075088799, + 0.006361931096762419, + 0.00792159978300333, + 0.008120953105390072, + 0.011822648346424103, + 0.006645249202847481, + 0.004208724945783615, + 0.0028883914928883314, + 0.004904822912067175, + 0.00799073837697506, + 0.004512427840381861, + 0.005761157721281052, + 0.005018836818635464, + 0.0052656140178442, + 0.01167099829763174, + 0.014776387251913548, + 0.008055360056459904, + 0.013408979400992393, + 0.0216697808355093, + 0.02675637975335121, + 0.027478668838739395, + 0.017088789492845535, + 0.00828642025589943, + 0.008642986416816711, + 0.007101209834218025, + 0.01209726370871067, + 0.012066694907844067, + 0.014932083897292614, + 0.015866294503211975, + 0.012534488923847675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.09975709766149521 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026382189244031906 + ], + "opposite_logprob_tokens": [ + -3.6513822078704834 + ], + "logprob_reward": -0.026382189244031906, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026382189244031906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16826151311397552, + 0.06422244757413864, + 0.08028662204742432, + 0.13287711143493652, + 0.04412401467561722, + 0.17966660857200623, + 0.07828187197446823, + 0.20780755579471588, + 0.026661472395062447, + 0.017810754477977753 + ], + "credit_map": [ + 0.021176809445023537, + 0.019692298024892807, + 0.024307815358042717, + 0.040577005594968796, + 0.01081065833568573, + 0.017646707594394684, + 0.018627194687724113, + 0.015423032455146313, + 0.01450915914028883, + 0.005047020502388477, + 0.007701966445893049, + 0.004672508221119642, + 0.011412633582949638, + 0.008827103301882744, + 0.0057651326060295105, + 0.00628692377358675, + 0.010101829655468464, + 0.009764285758137703, + 0.008654234930872917, + 0.00986857432872057, + 0.00736666563898325, + 0.015214246697723866, + 0.014215664938092232, + 0.005101121496409178, + 0.016200216487050056, + 0.009286654181778431, + 0.0186950471252203, + 0.010766047984361649, + 0.013680830597877502, + 0.0204781461507082, + 0.02709951251745224, + 0.016670653596520424, + 0.004981933161616325, + 0.004239370115101337, + 0.00533528346568346, + 0.003906935453414917, + 0.004496825393289328, + 0.00940605066716671, + 0.005498341750353575, + 0.006259276065975428, + 0.014369184151291847, + 0.009521336294710636, + 0.014864103868603706, + 0.014366047456860542, + 0.025195598602294922, + 0.044129639863967896, + 0.029348088428378105, + 0.027872612699866295, + 0.0056996578350663185, + 0.004057396203279495, + 0.007275703828781843, + 0.008017252199351788, + 0.008586838841438293, + 0.015557186678051949, + 0.017146585509181023, + 0.011941253207623959, + 0.018383439630270004, + 0.0150845255702734, + 0.030884182080626488, + 0.0555676594376564, + 0.01317523792386055, + 0.02109270915389061, + 0.019030822440981865, + 0.03458898514509201, + 0.0026409607380628586, + 0.0030393723864108324, + 0.0022470431867986917, + 0.0033975231926888227, + 0.003695449559018016, + 0.0048548090271651745, + 0.0036441260017454624, + 0.0031421866733580828, + 0.0016022089403122663, + 0.0017953830538317561, + 0.0014437122736126184, + 0.0021400251425802708, + 0.002486118348315358, + 0.003124244511127472, + 0.0029186333995312452, + 0.002300427295267582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.026382189244031906 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087356135249138 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087356135249138, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087356135249138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24793574213981628, + 0.09009746462106705, + 0.2030109167098999, + 0.1434887945652008, + 0.03905254602432251, + 0.09533921629190445, + 0.03950893133878708, + 0.045755572617053986, + 0.06493687629699707, + 0.030873974785208702 + ], + "credit_map": [ + 0.04538879543542862, + 0.031103068962693214, + 0.027918793261051178, + 0.046733129769563675, + 0.028330324217677116, + 0.019136516377329826, + 0.02892635576426983, + 0.020398758351802826, + 0.016204645857214928, + 0.010526282712817192, + 0.016890963539481163, + 0.007452469784766436, + 0.01526225171983242, + 0.00891067460179329, + 0.007632036227732897, + 0.007218134589493275, + 0.023119889199733734, + 0.020869052037596703, + 0.019320249557495117, + 0.018894782289862633, + 0.019353700801730156, + 0.037215180695056915, + 0.04770627245306969, + 0.01653176359832287, + 0.014246582984924316, + 0.011642961762845516, + 0.017619077116250992, + 0.012631784193217754, + 0.014709506183862686, + 0.02236388809978962, + 0.03394576162099838, + 0.01632922887802124, + 0.004488739185035229, + 0.0049609034322202206, + 0.004037821665406227, + 0.004555846098810434, + 0.005997481755912304, + 0.006043612957000732, + 0.004094718489795923, + 0.0048734224401414394, + 0.008836036548018456, + 0.007981806993484497, + 0.011631196364760399, + 0.007722754031419754, + 0.01221146248281002, + 0.020249303430318832, + 0.013901597820222378, + 0.01280504371970892, + 0.004541528411209583, + 0.0037181221414357424, + 0.004868776071816683, + 0.005073478911072016, + 0.004970587324351072, + 0.004998387768864632, + 0.005886639468371868, + 0.005451406817883253, + 0.004150536842644215, + 0.0029737704899162054, + 0.006631884723901749, + 0.006429978180676699, + 0.004502198658883572, + 0.008010964840650558, + 0.006891144905239344, + 0.006165089551359415, + 0.0053383405320346355, + 0.006776927039027214, + 0.0038968506269156933, + 0.005302745848894119, + 0.01113290898501873, + 0.014508807100355625, + 0.009999514557421207, + 0.007980775088071823, + 0.004319298546761274, + 0.0036072286311537027, + 0.005153091624379158, + 0.0027476870454847813, + 0.0027766490820795298, + 0.004159644246101379, + 0.0044266777113080025, + 0.003683696035295725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.011087356135249138 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041038114577531815 + ], + "opposite_logprob_tokens": [ + -5.504103660583496 + ], + "logprob_reward": -0.0041038114577531815, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041038114577531815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20993420481681824, + 0.11024545133113861, + 0.1726229190826416, + 0.14460048079490662, + 0.05744252726435661, + 0.060969144105911255, + 0.09948116540908813, + 0.04567373916506767, + 0.0615721121430397, + 0.03745824471116066 + ], + "credit_map": [ + 0.029537729918956757, + 0.0319860503077507, + 0.04794847220182419, + 0.028177130967378616, + 0.018200170248746872, + 0.011726859025657177, + 0.028011728078126907, + 0.014346063137054443, + 0.020322609692811966, + 0.015046288259327412, + 0.013209288008511066, + 0.01195554155856371, + 0.017290061339735985, + 0.013020752929151058, + 0.010135989636182785, + 0.009264920838177204, + 0.01496875286102295, + 0.02406826801598072, + 0.020794222131371498, + 0.014077365398406982, + 0.018180392682552338, + 0.018272407352924347, + 0.04257286712527275, + 0.019688626751303673, + 0.016028346493840218, + 0.01778358593583107, + 0.02374763786792755, + 0.013131041079759598, + 0.014727837406098843, + 0.021466946229338646, + 0.023531246930360794, + 0.014183856546878815, + 0.008283053524792194, + 0.006424793507903814, + 0.007778757717460394, + 0.007193689234554768, + 0.0071691833436489105, + 0.0076803481206297874, + 0.007415588013827801, + 0.00549711287021637, + 0.008319657295942307, + 0.007010926492512226, + 0.005319937132298946, + 0.00780853908509016, + 0.008320051245391369, + 0.008890212513506413, + 0.00923391617834568, + 0.006065903697162867, + 0.013397946022450924, + 0.012265868484973907, + 0.009994428604841232, + 0.01330379769206047, + 0.016180474311113358, + 0.00991151574999094, + 0.01392727717757225, + 0.010499854572117329, + 0.004809094592928886, + 0.004153323359787464, + 0.005720009095966816, + 0.007932011038064957, + 0.003765205154195428, + 0.007346196100115776, + 0.005338794551789761, + 0.006609104573726654, + 0.006222025491297245, + 0.005887218285351992, + 0.0073186433874070644, + 0.009452165104448795, + 0.007170775439590216, + 0.009228912182152271, + 0.00972476601600647, + 0.006567604374140501, + 0.0036270730197429657, + 0.004039565101265907, + 0.004107771907001734, + 0.003438766812905669, + 0.005212601274251938, + 0.006904149893671274, + 0.006293985061347485, + 0.0038343307096511126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0041038114577531815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306260123848915 + ], + "opposite_logprob_tokens": [ + -3.168062686920166 + ], + "logprob_reward": -0.04306260123848915, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306260123848915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1461106836795807, + 0.08263421803712845, + 0.14847145974636078, + 0.06787437200546265, + 0.05282626673579216, + 0.13120006024837494, + 0.0655062198638916, + 0.23076994717121124, + 0.05406349152326584, + 0.020543307065963745 + ], + "credit_map": [ + 0.019196122884750366, + 0.018466193228960037, + 0.028763432055711746, + 0.026326196268200874, + 0.01495729573071003, + 0.011928119696676731, + 0.014204851351678371, + 0.01226846594363451, + 0.0227335374802351, + 0.0057588121853768826, + 0.01597309298813343, + 0.005093549843877554, + 0.011511880904436111, + 0.008599222637712955, + 0.007162394933402538, + 0.005801723338663578, + 0.019651541486382484, + 0.016150183975696564, + 0.018079882487654686, + 0.015305251814424992, + 0.012089396826922894, + 0.014273655600845814, + 0.038556650280952454, + 0.014364907518029213, + 0.006981619168072939, + 0.004839539527893066, + 0.008773193694651127, + 0.006343302316963673, + 0.006768068298697472, + 0.011085486970841885, + 0.01682855375111103, + 0.006254606414586306, + 0.006137109827250242, + 0.004512179642915726, + 0.007622795645147562, + 0.006570028141140938, + 0.005489245522767305, + 0.006890700664371252, + 0.008162313140928745, + 0.007441896479576826, + 0.009163547307252884, + 0.007542584091424942, + 0.01613866724073887, + 0.008904870599508286, + 0.018322566524147987, + 0.030236266553401947, + 0.020949482917785645, + 0.01994207315146923, + 0.005964167881757021, + 0.0047296988777816296, + 0.0059941792860627174, + 0.007094746455550194, + 0.009893829934298992, + 0.012410693801939487, + 0.01071193441748619, + 0.008706972934305668, + 0.01995699852705002, + 0.016926990821957588, + 0.032821182161569595, + 0.0644579827785492, + 0.017465077340602875, + 0.019183829426765442, + 0.024734098464250565, + 0.03522379696369171, + 0.004817140754312277, + 0.004971889313310385, + 0.0039042001590132713, + 0.0060064587742090225, + 0.007748361676931381, + 0.012485644780099392, + 0.007094570901244879, + 0.007035224698483944, + 0.002538583241403103, + 0.0024639435578137636, + 0.0023212111555039883, + 0.0023114655632525682, + 0.0023717922158539295, + 0.002413556445389986, + 0.003785389242693782, + 0.0023373663425445557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.04306260123848915 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic, low-angle, side-on tracking shot follows an elegant black Maybach as it speeds through a rain-soaked city street at night. The camera's focus is pinpointed on the vehicle's front wheels, their high-gloss multi-spoke alloy wheels spinning at high speed in a blur of dynamic light and shadow. The iconic logo embedded in their centers occasionally gleams coldly under the passing streetlights. The wet black tires slice through the puddle at breakneck speed, violently flinging a vast, spectacular fan-shaped curtain of water to the rear and sides. Countless shimmering droplets, momentarily illuminated by the headlights and the shimmering neon lights in the background, glitter like diamonds tossed into the air. In the background, the wet asphalt reflects the cold white light of the streetlights and the iridescent billboards, creating distorted, flowing reflections. The image, dominated by deep black and wet asphalt gray, is punctuated by neon magenta and electric blue, creating a cool, urban atmosphere imbued with power, speed, and luxury.", + "seed": 6315, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_2/index_3081_seed_6315_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017049228865653278, + "vlm_logprob": -0.017049228865653278, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.017049228865653278, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047654714435338974 + ], + "opposite_logprob_tokens": [ + -3.172654628753662 + ], + "logprob_reward": -0.047654714435338974, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.047654714435338974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06053268909454346, + 0.07467833161354065, + 0.04126990959048271, + 0.22810763120651245, + 0.25291335582733154, + 0.08085691183805466, + 0.10149113088846207, + 0.0997520387172699, + 0.037362344563007355, + 0.02303568460047245 + ], + "credit_map": [ + 0.006278985645622015, + 0.009658721275627613, + 0.006481464020907879, + 0.005107001401484013, + 0.010475756600499153, + 0.008735816925764084, + 0.007964914664626122, + 0.005830032750964165, + 0.007301224861294031, + 0.02144908718764782, + 0.011226313188672066, + 0.006439140997827053, + 0.004546123556792736, + 0.004574154503643513, + 0.00477433018386364, + 0.014367957599461079, + 0.00431759562343359, + 0.004895419348031282, + 0.004167346283793449, + 0.0026668605860322714, + 0.005395878572016954, + 0.00905043538659811, + 0.006318556610494852, + 0.0044578188098967075, + 0.031279556453228, + 0.014800834469497204, + 0.011385980993509293, + 0.012065056711435318, + 0.008824510499835014, + 0.058260440826416016, + 0.01242415513843298, + 0.07906709611415863, + 0.04099295660853386, + 0.020399436354637146, + 0.030113086104393005, + 0.01574057526886463, + 0.014956917613744736, + 0.07333593815565109, + 0.02624615840613842, + 0.031128283590078354, + 0.012002869509160519, + 0.00906357541680336, + 0.009081121534109116, + 0.009325806051492691, + 0.008665462955832481, + 0.01450891699641943, + 0.00778912054374814, + 0.010420038364827633, + 0.014483984559774399, + 0.012924300506711006, + 0.006866001524031162, + 0.010408639907836914, + 0.013299882411956787, + 0.019733212888240814, + 0.010765323415398598, + 0.013009787537157536, + 0.012513229623436928, + 0.007842023856937885, + 0.007522277534008026, + 0.009117078967392445, + 0.013684970326721668, + 0.023409295827150345, + 0.014150355011224747, + 0.01151280477643013, + 0.007221871986985207, + 0.00298303016461432, + 0.003297799965366721, + 0.004021222237497568, + 0.004103681072592735, + 0.0038895942270755768, + 0.0076744756661355495, + 0.004170668311417103, + 0.00212182872928679, + 0.00176757771987468, + 0.0027125398628413677, + 0.0030508076306432486, + 0.002704836195334792, + 0.0033920102287083864, + 0.004454108886420727, + 0.0028319754637777805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.047654714435338974 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007648232858628035 + ], + "opposite_logprob_tokens": [ + -4.882648468017578 + ], + "logprob_reward": -0.007648232858628035, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007648232858628035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3453814685344696, + 0.08107653260231018, + 0.062215592712163925, + 0.06805822253227234, + 0.10925664752721786, + 0.061511557549238205, + 0.1128709688782692, + 0.04730064049363136, + 0.07945740967988968, + 0.03287104517221451 + ], + "credit_map": [ + 0.021101579070091248, + 0.03142683207988739, + 0.027317630127072334, + 0.027911946177482605, + 0.06448034942150116, + 0.05081338435411453, + 0.08285900950431824, + 0.03947073966264725, + 0.008177005685865879, + 0.027728211134672165, + 0.01244147028774023, + 0.004408326931297779, + 0.0037972964346408844, + 0.003877531737089157, + 0.003741159802302718, + 0.016905521973967552, + 0.006040979642421007, + 0.00583234429359436, + 0.006982824765145779, + 0.00404708506539464, + 0.006698691286146641, + 0.010845712386071682, + 0.015370001085102558, + 0.006397954188287258, + 0.007212386932224035, + 0.0045577725395560265, + 0.003863481106236577, + 0.0057969833724200726, + 0.0064872936345636845, + 0.016734080389142036, + 0.009576618671417236, + 0.013829600997269154, + 0.014617358334362507, + 0.007374433800578117, + 0.009235046803951263, + 0.007941336371004581, + 0.008724291808903217, + 0.03379096835851669, + 0.012934757396578789, + 0.01463845744729042, + 0.006857775617390871, + 0.008013500832021236, + 0.007576802279800177, + 0.00756794074550271, + 0.006540989503264427, + 0.009064779616892338, + 0.008593142963945866, + 0.007296629715710878, + 0.017533665522933006, + 0.009091932326555252, + 0.008784538134932518, + 0.010404800996184349, + 0.015905320644378662, + 0.022228149697184563, + 0.0135669419541955, + 0.015355623327195644, + 0.005128673277795315, + 0.004547628108412027, + 0.003294586669653654, + 0.004008261486887932, + 0.007552327122539282, + 0.011790558695793152, + 0.0054945875890553, + 0.005484017543494701, + 0.009749507531523705, + 0.005263371393084526, + 0.004791792947798967, + 0.01031676959246397, + 0.011848144233226776, + 0.01267324760556221, + 0.017874646931886673, + 0.006939929910004139, + 0.0030800800304859877, + 0.0032708081416785717, + 0.004346505738794804, + 0.003453838638961315, + 0.003915227018296719, + 0.004095683805644512, + 0.006437597330659628, + 0.004271300975233316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.007648232858628035 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791087359189987 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009791087359189987, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791087359189987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08621779084205627, + 0.08244367688894272, + 0.05375634506344795, + 0.04763112962245941, + 0.1331280618906021, + 0.15144437551498413, + 0.27232539653778076, + 0.11488541960716248, + 0.03592894971370697, + 0.022238807752728462 + ], + "credit_map": [ + 0.006343896966427565, + 0.010951176285743713, + 0.007143361959606409, + 0.006818892899900675, + 0.011842960491776466, + 0.014819751493632793, + 0.016431406140327454, + 0.011866348795592785, + 0.009102101437747478, + 0.028926195576786995, + 0.010295757092535496, + 0.0033393593039363623, + 0.0032747616060078144, + 0.004270230419933796, + 0.0043825386092066765, + 0.018852729350328445, + 0.006257949862629175, + 0.005930318031460047, + 0.007336800917983055, + 0.004138675518333912, + 0.006674309726804495, + 0.008549303747713566, + 0.007778354454785585, + 0.0070906346663832664, + 0.0078012216836214066, + 0.006113238632678986, + 0.0027976937126368284, + 0.005124680697917938, + 0.0049306852743029594, + 0.006191067863255739, + 0.004162264987826347, + 0.010510275140404701, + 0.018168320879340172, + 0.008753395639359951, + 0.019053615629673004, + 0.009342346340417862, + 0.008402765728533268, + 0.038315802812576294, + 0.013974905014038086, + 0.01711690053343773, + 0.01664283685386181, + 0.015970170497894287, + 0.019084906205534935, + 0.012997513636946678, + 0.013655978254973888, + 0.03401130810379982, + 0.017186135053634644, + 0.021895522251725197, + 0.040377210825681686, + 0.028627732768654823, + 0.019368359819054604, + 0.027284199371933937, + 0.03843263164162636, + 0.054443348199129105, + 0.026110250502824783, + 0.03768168389797211, + 0.011577233672142029, + 0.007635992951691151, + 0.009965960867702961, + 0.009810550138354301, + 0.014136440120637417, + 0.02335609868168831, + 0.024187415838241577, + 0.014215733855962753, + 0.003195014549419284, + 0.0028126707766205072, + 0.002468359423801303, + 0.004478032235056162, + 0.0053467522375285625, + 0.004233378451317549, + 0.009454264305531979, + 0.00394047936424613, + 0.001765460125170648, + 0.001743893837556243, + 0.003614113898947835, + 0.0030591574031859636, + 0.002188084414228797, + 0.003194797318428755, + 0.004135890398174524, + 0.002537410706281662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.009791087359189987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202939346432686 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014202939346432686, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202939346432686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058820731937885284, + 0.05441128462553024, + 0.04513362795114517, + 0.060930319130420685, + 0.19698049128055573, + 0.10177461057901382, + 0.42066630721092224, + 0.04292520880699158, + 0.011779521591961384, + 0.006577880121767521 + ], + "credit_map": [ + 0.0039354427717626095, + 0.0066807144321501255, + 0.006251111626625061, + 0.005105540156364441, + 0.011816572397947311, + 0.008189539425075054, + 0.010541326366364956, + 0.006300484761595726, + 0.0066941059194505215, + 0.013714016415178776, + 0.008293514139950275, + 0.0073234569281339645, + 0.0035244955215603113, + 0.0041038477793335915, + 0.00408730935305357, + 0.006670538801699877, + 0.005302937235683203, + 0.0062444331124424934, + 0.005436519160866737, + 0.003997130785137415, + 0.006392169743776321, + 0.005633930210024118, + 0.00706902239471674, + 0.005057486705482006, + 0.007862241007387638, + 0.003508404828608036, + 0.002715401817113161, + 0.0027012352366000414, + 0.0031675377395004034, + 0.016528965905308723, + 0.003631134983152151, + 0.020815398544073105, + 0.029361378401517868, + 0.013139955699443817, + 0.03355172276496887, + 0.011824380606412888, + 0.010725229047238827, + 0.05550245940685272, + 0.019048741087317467, + 0.023826617747545242, + 0.010685171000659466, + 0.011416474357247353, + 0.007406147196888924, + 0.008289760909974575, + 0.008205806836485863, + 0.024933286011219025, + 0.014096066355705261, + 0.016741901636123657, + 0.05630093812942505, + 0.045864835381507874, + 0.026781396940350533, + 0.04588548094034195, + 0.058855652809143066, + 0.08474358916282654, + 0.04451155290007591, + 0.05772289261221886, + 0.005955916829407215, + 0.004080676473677158, + 0.005654139444231987, + 0.005392454098910093, + 0.0048567564226686954, + 0.005903908982872963, + 0.006767979357391596, + 0.004313373938202858, + 0.0010133396135643125, + 0.0008681730250827968, + 0.000799975881818682, + 0.0012884901370853186, + 0.0019513670122250915, + 0.0016189520247280598, + 0.0027699714992195368, + 0.0014692521654069424, + 0.0006300363456830382, + 0.0006001809379085898, + 0.0009322749101556838, + 0.0008432338945567608, + 0.0007030487759038806, + 0.0008934881188906729, + 0.0010875830193981528, + 0.0008880339446477592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.014202939346432686 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949170328676701 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005949170328676701, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949170328676701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11768906563520432, + 0.026815813034772873, + 0.09505723416805267, + 0.18386094272136688, + 0.09630090743303299, + 0.18055325746536255, + 0.16932839155197144, + 0.0379377081990242, + 0.05731354281306267, + 0.0351431705057621 + ], + "credit_map": [ + 0.007812078110873699, + 0.01256615575402975, + 0.011595085263252258, + 0.012122062034904957, + 0.01814793236553669, + 0.012832075357437134, + 0.02881448157131672, + 0.01379918958991766, + 0.0028401007875800133, + 0.0038348701782524586, + 0.005132318940013647, + 0.002705223858356476, + 0.0023067332804203033, + 0.0036742070224136114, + 0.0033674791920930147, + 0.0029548807069659233, + 0.011489356867969036, + 0.006795299705117941, + 0.011357484385371208, + 0.00533870467916131, + 0.009112291969358921, + 0.012562843970954418, + 0.02840583026409149, + 0.009995422326028347, + 0.0210382379591465, + 0.013340190052986145, + 0.010175660252571106, + 0.016112636774778366, + 0.023536011576652527, + 0.04128044471144676, + 0.02627149224281311, + 0.032106272876262665, + 0.013740316033363342, + 0.007815660908818245, + 0.01584051549434662, + 0.005865882150828838, + 0.010505625046789646, + 0.019042985513806343, + 0.010087925009429455, + 0.013402000069618225, + 0.011970974504947662, + 0.011451620608568192, + 0.010668625123798847, + 0.012209533713757992, + 0.013866779394447803, + 0.057495880872011185, + 0.026508331298828125, + 0.036381516605615616, + 0.02524682879447937, + 0.016369163990020752, + 0.01254219375550747, + 0.01666041649878025, + 0.02392892725765705, + 0.03392630070447922, + 0.017107825726270676, + 0.023546742275357246, + 0.0035273521207273006, + 0.0031119096092879772, + 0.0033226204104721546, + 0.004006081726402044, + 0.005248445086181164, + 0.008640234358608723, + 0.00565282441675663, + 0.004428240470588207, + 0.009159650653600693, + 0.004788192920386791, + 0.006427253130823374, + 0.006212977226823568, + 0.0065196300856769085, + 0.00843345932662487, + 0.009726478718221188, + 0.006045902613550425, + 0.0030148234218358994, + 0.0029335434082895517, + 0.005545560270547867, + 0.003917396068572998, + 0.003925136756151915, + 0.003550897818058729, + 0.007527281064540148, + 0.0047285291366279125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.005949170328676701 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic, low-angle, side-on tracking shot follows an elegant black Maybach as it speeds through a rain-soaked city street at night. The camera's focus is pinpointed on the vehicle's front wheels, their high-gloss multi-spoke alloy wheels spinning at high speed in a blur of dynamic light and shadow. The iconic logo embedded in their centers occasionally gleams coldly under the passing streetlights. The wet black tires slice through the puddle at breakneck speed, violently flinging a vast, spectacular fan-shaped curtain of water to the rear and sides. Countless shimmering droplets, momentarily illuminated by the headlights and the shimmering neon lights in the background, glitter like diamonds tossed into the air. In the background, the wet asphalt reflects the cold white light of the streetlights and the iridescent billboards, creating distorted, flowing reflections. The image, dominated by deep black and wet asphalt gray, is punctuated by neon magenta and electric blue, creating a cool, urban atmosphere imbued with power, speed, and luxury.", + "seed": 6316, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_2/index_3081_seed_6316_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02359365448355675, + "vlm_logprob": -0.02359365448355675, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02359365448355675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07277648150920868 + ], + "opposite_logprob_tokens": [ + -2.6977765560150146 + ], + "logprob_reward": -0.07277648150920868, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07277648150920868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22306811809539795, + 0.1518414169549942, + 0.16659802198410034, + 0.04175984114408493, + 0.08450441807508469, + 0.03571496158838272, + 0.05220474675297737, + 0.04136410355567932, + 0.10792451351881027, + 0.0950198546051979 + ], + "credit_map": [ + 0.02433173544704914, + 0.021804310381412506, + 0.06931770592927933, + 0.06243603676557541, + 0.010430434718728065, + 0.010491594672203064, + 0.012626800686120987, + 0.011629490181803703, + 0.018085123971104622, + 0.013138966634869576, + 0.03538694605231285, + 0.027746085077524185, + 0.016773324459791183, + 0.011109728366136551, + 0.018257983028888702, + 0.011343268677592278, + 0.015372397378087044, + 0.022838490083813667, + 0.015751127153635025, + 0.012315243482589722, + 0.039644431322813034, + 0.017924202606081963, + 0.026182085275650024, + 0.01657002978026867, + 0.004260476212948561, + 0.004597232211381197, + 0.004209232050925493, + 0.0063480292446911335, + 0.006162215489894152, + 0.00529179023578763, + 0.004991085734218359, + 0.005899779032915831, + 0.006924040615558624, + 0.012964682653546333, + 0.009479440748691559, + 0.009032432921230793, + 0.017072541639208794, + 0.012181348167359829, + 0.009650168940424919, + 0.007199765648692846, + 0.004806329496204853, + 0.005876828916370869, + 0.0035331863909959793, + 0.004275988321751356, + 0.006475405767560005, + 0.0048866611905395985, + 0.0029217489063739777, + 0.002938810270279646, + 0.008039035834372044, + 0.007088611833751202, + 0.007700121030211449, + 0.003922224976122379, + 0.009434300474822521, + 0.004850167315453291, + 0.005317980889230967, + 0.005852305330336094, + 0.004835366737097502, + 0.011390850879251957, + 0.003337464528158307, + 0.0027446532621979713, + 0.002680862322449684, + 0.0053210933692753315, + 0.0033249547705054283, + 0.007728859316557646, + 0.01796182617545128, + 0.01178442221134901, + 0.016223711892962456, + 0.008238546550273895, + 0.01266628410667181, + 0.011567601934075356, + 0.01590472273528576, + 0.013577397912740707, + 0.011881590820848942, + 0.014766691252589226, + 0.010455306619405746, + 0.012804378755390644, + 0.009963552467525005, + 0.010700198821723461, + 0.01112178061157465, + 0.013326353393495083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.07277648150920868 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255783811211586 + ], + "opposite_logprob_tokens": [ + -4.3875579833984375 + ], + "logprob_reward": -0.01255783811211586, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255783811211586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12482482194900513, + 0.06818705052137375, + 0.26437363028526306, + 0.04545844346284866, + 0.02411678060889244, + 0.030960744246840477, + 0.07515377551317215, + 0.18827404081821442, + 0.10554271191358566, + 0.07310798019170761 + ], + "credit_map": [ + 0.013708817772567272, + 0.012919721193611622, + 0.035285584628582, + 0.023444868624210358, + 0.013156048022210598, + 0.008077944628894329, + 0.00901890080422163, + 0.009212932549417019, + 0.009733736515045166, + 0.006549615412950516, + 0.011460278183221817, + 0.014140209183096886, + 0.007630971726030111, + 0.005261002574115992, + 0.008787560276687145, + 0.004623674787580967, + 0.03313688188791275, + 0.01894872635602951, + 0.02865302376449108, + 0.027234654873609543, + 0.02904593199491501, + 0.024407055228948593, + 0.05882697179913521, + 0.04412039741873741, + 0.00444452092051506, + 0.004207403399050236, + 0.003621115582063794, + 0.006050724536180496, + 0.00714441854506731, + 0.010145037434995174, + 0.003893424989655614, + 0.005951800849288702, + 0.0024753245525062084, + 0.003104674629867077, + 0.0033999301958829165, + 0.0024915365502238274, + 0.0029389536939561367, + 0.0028588874265551567, + 0.004465988837182522, + 0.0023814851883798838, + 0.0030079644639045, + 0.006468358915299177, + 0.0031746672466397285, + 0.0030030105262994766, + 0.004854022525250912, + 0.004652538802474737, + 0.002735674846917391, + 0.0030645085498690605, + 0.012338840402662754, + 0.01001431979238987, + 0.010882253758609295, + 0.008025947958230972, + 0.010500284843146801, + 0.008547963574528694, + 0.008378938771784306, + 0.006465231068432331, + 0.04151858389377594, + 0.06706152856349945, + 0.013052234426140785, + 0.012984782457351685, + 0.011217186227440834, + 0.025246554985642433, + 0.009318268857896328, + 0.007874904200434685, + 0.013284806162118912, + 0.010968499816954136, + 0.02020704373717308, + 0.008671274408698082, + 0.012200434692203999, + 0.007524095941334963, + 0.01443617045879364, + 0.01825038716197014, + 0.010623226873576641, + 0.01347148697823286, + 0.008150803856551647, + 0.010832424275577068, + 0.007606884930282831, + 0.007450681645423174, + 0.007203094661235809, + 0.007769374176859856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.01255783811211586 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979037955403328 + ], + "opposite_logprob_tokens": [ + -4.634790420532227 + ], + "logprob_reward": -0.00979037955403328, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979037955403328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0646739974617958, + 0.18044809997081757, + 0.11109854280948639, + 0.2264920026063919, + 0.07290436327457428, + 0.03382641077041626, + 0.05989118292927742, + 0.11778313666582108, + 0.08758632093667984, + 0.04529586806893349 + ], + "credit_map": [ + 0.008849975652992725, + 0.01202483382076025, + 0.01020465325564146, + 0.008642840199172497, + 0.007376078516244888, + 0.005916931666433811, + 0.006385200656950474, + 0.005273476708680391, + 0.012227999046444893, + 0.012185534462332726, + 0.01820177026093006, + 0.02506352588534355, + 0.01972794160246849, + 0.023782018572092056, + 0.042028676718473434, + 0.027230646461248398, + 0.012905190698802471, + 0.016032928600907326, + 0.010626425966620445, + 0.008935466408729553, + 0.018664605915546417, + 0.01173184160143137, + 0.01627703569829464, + 0.015925047919154167, + 0.021777966991066933, + 0.031157853081822395, + 0.02327302098274231, + 0.042454373091459274, + 0.02727249637246132, + 0.025188427418470383, + 0.030019670724868774, + 0.02534818835556507, + 0.008239345625042915, + 0.010261754505336285, + 0.011823225766420364, + 0.008220303803682327, + 0.009462298825383186, + 0.008198811672627926, + 0.009131821803748608, + 0.007566806394606829, + 0.0043959105387330055, + 0.004828507546335459, + 0.004482941702008247, + 0.0031699487008154392, + 0.004851812496781349, + 0.005092262756079435, + 0.003458153223618865, + 0.0035468752030283213, + 0.007049533538520336, + 0.007701292168349028, + 0.00694834953173995, + 0.007884510792791843, + 0.01139776036143303, + 0.0051777055487036705, + 0.007908202707767487, + 0.0058238268829882145, + 0.01745653711259365, + 0.0379469208419323, + 0.006583221256732941, + 0.008205961436033249, + 0.009737350046634674, + 0.01592508889734745, + 0.007424085400998592, + 0.01450397353619337, + 0.018239181488752365, + 0.011572561226785183, + 0.015150114893913269, + 0.008148693479597569, + 0.010008463636040688, + 0.009358472190797329, + 0.007976134307682514, + 0.007132698316127062, + 0.006313075311481953, + 0.006091808434575796, + 0.003698736662045121, + 0.005189984571188688, + 0.004326654132455587, + 0.005623127333819866, + 0.00674480153247714, + 0.007307678461074829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.00979037955403328 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420411467552185 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.01420411467552185, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420411467552185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11416824907064438, + 0.307586133480072, + 0.08029886335134506, + 0.11846187710762024, + 0.049432333558797836, + 0.046184439212083817, + 0.0846058651804924, + 0.057855162769556046, + 0.1111195906996727, + 0.03028746508061886 + ], + "credit_map": [ + 0.012233264744281769, + 0.0173196978867054, + 0.024683458730578423, + 0.02313297428190708, + 0.010467327199876308, + 0.008326181210577488, + 0.01105696614831686, + 0.006948390044271946, + 0.025047002360224724, + 0.021659839898347855, + 0.04087020084261894, + 0.045449141412973404, + 0.02770223654806614, + 0.02870282717049122, + 0.07673289626836777, + 0.04142199456691742, + 0.009014150127768517, + 0.00994082260876894, + 0.007882365956902504, + 0.006922861095517874, + 0.011708544567227364, + 0.00797681137919426, + 0.013785656541585922, + 0.013067654334008694, + 0.012139332480728626, + 0.008445280604064465, + 0.018620438873767853, + 0.008439072407782078, + 0.02843317575752735, + 0.01822701282799244, + 0.009888344444334507, + 0.014269223436713219, + 0.004444520454853773, + 0.006796885281801224, + 0.006713181734085083, + 0.004860355984419584, + 0.009203832596540451, + 0.007490762509405613, + 0.005221690982580185, + 0.004701106809079647, + 0.004977978300303221, + 0.006929880939424038, + 0.003507679793983698, + 0.0044835349544882774, + 0.009901992976665497, + 0.007984234020113945, + 0.00430951127782464, + 0.004089627414941788, + 0.00941134337335825, + 0.010626567527651787, + 0.009783984161913395, + 0.01020783744752407, + 0.013503789901733398, + 0.010367225855588913, + 0.01147657074034214, + 0.00922855269163847, + 0.009597636759281158, + 0.019060514867305756, + 0.004065440036356449, + 0.004630391485989094, + 0.004141607787460089, + 0.008636998012661934, + 0.003285310696810484, + 0.004437267780303955, + 0.020153431221842766, + 0.010729365050792694, + 0.02215586043894291, + 0.010674752295017242, + 0.012291439808905125, + 0.012972438707947731, + 0.009759933687746525, + 0.012382376939058304, + 0.004534329287707806, + 0.00549250328913331, + 0.0035691503435373306, + 0.004251485224813223, + 0.003252391004934907, + 0.00297721428796649, + 0.0032857428304851055, + 0.0029246520716696978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.01420411467552185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elegant black Maybach?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a rain-soaked city street at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera shot taken from a low angle in a side-on tracking position?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera focus pinpointed on the vehicle's front wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car speed through puddles, flinging a fan-shaped curtain of water to the rear and sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639458566904068 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639458566904068, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639458566904068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3439316749572754, + 0.12177103757858276, + 0.07516736537218094, + 0.0772971659898758, + 0.040052298456430435, + 0.014212292619049549, + 0.08438318222761154, + 0.07217253744602203, + 0.10196338593959808, + 0.06904906034469604 + ], + "credit_map": [ + 0.030659427866339684, + 0.027072224766016006, + 0.11808229982852936, + 0.10672997683286667, + 0.022179076448082924, + 0.012757090851664543, + 0.01560945250093937, + 0.010842105373740196, + 0.01326918974518776, + 0.007657994050532579, + 0.023217761889100075, + 0.020160870626568794, + 0.011867097578942776, + 0.008727672509849072, + 0.024968290701508522, + 0.011902161873877048, + 0.006289041135460138, + 0.006343672052025795, + 0.007137068547308445, + 0.006362260319292545, + 0.012863719835877419, + 0.00945682916790247, + 0.013537880033254623, + 0.013176893815398216, + 0.007740319706499577, + 0.004837514366954565, + 0.01153064239770174, + 0.005587791558355093, + 0.02145734615623951, + 0.010524315759539604, + 0.006038586609065533, + 0.009580651298165321, + 0.006463715806603432, + 0.005679194815456867, + 0.004517997615039349, + 0.0036197409499436617, + 0.0065105087123811245, + 0.0040312595665454865, + 0.005555257201194763, + 0.003674625651910901, + 0.0016039141919463873, + 0.002035845536738634, + 0.0017858027713373303, + 0.0018915040418505669, + 0.0018622282659634948, + 0.0017405095277354121, + 0.0014047276927158237, + 0.0018877605907619, + 0.009331907145678997, + 0.010040892288088799, + 0.012534910812973976, + 0.0073366849683225155, + 0.013858526013791561, + 0.01257993932813406, + 0.008735110983252525, + 0.009965211153030396, + 0.013131888583302498, + 0.024602403864264488, + 0.0040167514234781265, + 0.005932417698204517, + 0.004983330145478249, + 0.010458080098032951, + 0.003730052849277854, + 0.005317602772265673, + 0.00947994738817215, + 0.0059660449624061584, + 0.022398145869374275, + 0.012667231261730194, + 0.011871503666043282, + 0.0074152396991848946, + 0.011575760319828987, + 0.020589502528309822, + 0.007748574484139681, + 0.011458145454525948, + 0.0077739981934428215, + 0.00849265605211258, + 0.006973422132432461, + 0.007807983085513115, + 0.007973226718604565, + 0.010821053758263588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.008639458566904068 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_3/index_tensor([1465, 1465])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_3/index_tensor([1465, 1465])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..76f6d41ec6d2aa185342c513bb68baefbbb53bfa --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_3/index_tensor([1465, 1465])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, medium shot takes us to a deserted subway platform late at night. An elderly street performer takes center stage, wearing a well-worn, dark fedora, his graying hair tangled from beneath its brim. His face is lined with wrinkles as deep as a map. Clad in a heavy, dark gray wool coat with frayed edges, he closes his eyes in concentration, completely absorbed in the brass alto saxophone he holds. A few pale fluorescent lights on the ceiling provide the sole source of light, their cold, vertical beams illuminating his cheeks, which puff slightly from his intense playing, and casting a solitary, elongated shadow on the white tiled wall behind him. His fingers dance skillfully and briskly across the mother-of-pearl keys, warmed by constant use, his body swaying gently to the melancholic melody he produces. The melodious and slightly sad saxophone music echoes clearly in the empty underground space, interweaving with the faint sound of wind coming from the distant tunnel. The whole scene, set against the backdrop of cold colors, creates a quiet atmosphere of extreme loneliness yet full of artistic resilience.", + "seed": 10758, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_3/index_1465_seed_10758_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.042463130317628384, + "vlm_logprob": -0.042463130317628384, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.042463130317628384, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11776957660913467 + ], + "opposite_logprob_tokens": [ + -2.367769479751587 + ], + "logprob_reward": -0.11776957660913467, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11776957660913467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2323530614376068, + 0.1101645678281784, + 0.14202645421028137, + 0.08883132040500641, + 0.12046954780817032, + 0.1337159126996994, + 0.03870207443833351, + 0.04345143213868141, + 0.06954799592494965, + 0.020737633109092712 + ], + "credit_map": [ + 0.02455969899892807, + 0.026952635496854782, + 0.05210573971271515, + 0.025067733600735664, + 0.017450610175728798, + 0.018145643174648285, + 0.030595235526561737, + 0.03747575357556343, + 0.01566684991121292, + 0.009346676990389824, + 0.02422017604112625, + 0.009560683742165565, + 0.011065931059420109, + 0.01234920509159565, + 0.0135825015604496, + 0.014372547157108784, + 0.02424011379480362, + 0.023573385551571846, + 0.0252241138368845, + 0.01469619944691658, + 0.011653540655970573, + 0.010866060853004456, + 0.013630978763103485, + 0.01814206875860691, + 0.008094769902527332, + 0.008710920810699463, + 0.023587536066770554, + 0.010943815112113953, + 0.008999551646411419, + 0.0081624835729599, + 0.010193760506808758, + 0.010138478130102158, + 0.011129556223750114, + 0.017739415168762207, + 0.02798624522984028, + 0.017486410215497017, + 0.010249095037579536, + 0.013693553395569324, + 0.011469367891550064, + 0.010715903714299202, + 0.011239309795200825, + 0.015318130142986774, + 0.02025710791349411, + 0.01673838123679161, + 0.01388213224709034, + 0.011973019689321518, + 0.01726614497601986, + 0.027041688561439514, + 0.0026153430808335543, + 0.004884387832134962, + 0.004165980964899063, + 0.003953726030886173, + 0.003872525179758668, + 0.009625035338103771, + 0.0037886027712374926, + 0.005796472076326609, + 0.006484032142907381, + 0.003298442345112562, + 0.010539785027503967, + 0.003991583362221718, + 0.0036760433577001095, + 0.0029492515604943037, + 0.005878404248505831, + 0.00663389079272747, + 0.012401962652802467, + 0.007291547488421202, + 0.015727359801530838, + 0.008892890065908432, + 0.004422299563884735, + 0.005231162998825312, + 0.007618962321430445, + 0.007961814291775227, + 0.002178920665755868, + 0.002272817073389888, + 0.00440049497410655, + 0.0025497805327177048, + 0.002441414399072528, + 0.001861294382251799, + 0.002427024533972144, + 0.0026058864314109087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.11776957660913467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331940270960331 + ], + "opposite_logprob_tokens": [ + -3.7733194828033447 + ], + "logprob_reward": -0.02331940270960331, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331940270960331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3637353181838989, + 0.10686740279197693, + 0.08846566081047058, + 0.04033826291561127, + 0.050032127648591995, + 0.09789542853832245, + 0.13892576098442078, + 0.042275864630937576, + 0.024862870573997498, + 0.04660129174590111 + ], + "credit_map": [ + 0.021807625889778137, + 0.06327756494283676, + 0.051180753856897354, + 0.053414251655340195, + 0.04114789143204689, + 0.048525258898735046, + 0.044192805886268616, + 0.040189165621995926, + 0.017925070598721504, + 0.006812526378780603, + 0.018139060586690903, + 0.010949326679110527, + 0.009300931356847286, + 0.008733556605875492, + 0.015294556505978107, + 0.019712364301085472, + 0.009385088458657265, + 0.010027359239757061, + 0.014490879140794277, + 0.014210231602191925, + 0.009025361388921738, + 0.007313765585422516, + 0.011830276809632778, + 0.012182699516415596, + 0.005804488435387611, + 0.004190633073449135, + 0.0065255616791546345, + 0.0058399406261742115, + 0.004192054271697998, + 0.005119382403790951, + 0.00476967403665185, + 0.003896527923643589, + 0.003617827547714114, + 0.0060741775669157505, + 0.006197670940309763, + 0.007273922208696604, + 0.008945695124566555, + 0.005625404417514801, + 0.006175096146762371, + 0.00612233579158783, + 0.008762933313846588, + 0.012587997131049633, + 0.02345474436879158, + 0.00979357399046421, + 0.009963776916265488, + 0.009670472703874111, + 0.014067563228309155, + 0.009594363160431385, + 0.005042549222707748, + 0.01367273461073637, + 0.008165673352777958, + 0.00949148926883936, + 0.0046541267074644566, + 0.038910161703825, + 0.019792938604950905, + 0.03919610008597374, + 0.003385564312338829, + 0.0035597074311226606, + 0.007180951535701752, + 0.004213184118270874, + 0.005519185680896044, + 0.0047810813412070274, + 0.008491144515573978, + 0.005145044531673193, + 0.002204018412157893, + 0.0030541287269443274, + 0.0034200800582766533, + 0.004451780114322901, + 0.002489099046215415, + 0.002330891089513898, + 0.0038626138120889664, + 0.0030502600129693747, + 0.004352051764726639, + 0.006799756549298763, + 0.005700819659978151, + 0.007857756689190865, + 0.005667869001626968, + 0.005690091289579868, + 0.006039659027010202, + 0.004493287298828363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.02331940270960331 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371734917163849 + ], + "opposite_logprob_tokens": [ + -3.408717393875122 + ], + "logprob_reward": -0.03371734917163849, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371734917163849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26336556673049927, + 0.09354367107152939, + 0.11235883831977844, + 0.08075729012489319, + 0.06474801898002625, + 0.09208224713802338, + 0.19605614244937897, + 0.04575392231345177, + 0.024294335395097733, + 0.027040032669901848 + ], + "credit_map": [ + 0.01711489073932171, + 0.03289782628417015, + 0.045682765543460846, + 0.039025306701660156, + 0.031125977635383606, + 0.03626531735062599, + 0.03217800334095955, + 0.029075467959046364, + 0.009085433557629585, + 0.011478405445814133, + 0.010120837949216366, + 0.010792437009513378, + 0.014765053056180477, + 0.010560974478721619, + 0.014427797868847847, + 0.012312723323702812, + 0.010369433090090752, + 0.016318177804350853, + 0.014028958044946194, + 0.01660291664302349, + 0.014504948630928993, + 0.010941451415419579, + 0.015088779851794243, + 0.01450415886938572, + 0.009133913554251194, + 0.008914479054510593, + 0.017607374116778374, + 0.008293068036437035, + 0.008308768272399902, + 0.008557288907468319, + 0.011081477627158165, + 0.008860914967954159, + 0.0048471661284565926, + 0.007824166677892208, + 0.00976849626749754, + 0.008612244389951229, + 0.009766546078026295, + 0.009273244068026543, + 0.006823408883064985, + 0.00783274695277214, + 0.00848674587905407, + 0.011712558567523956, + 0.019419513642787933, + 0.01063667144626379, + 0.014549351297318935, + 0.007336994633078575, + 0.01165094692260027, + 0.008289458230137825, + 0.007964029908180237, + 0.02153853513300419, + 0.011426793411374092, + 0.015198851935565472, + 0.008559205569326878, + 0.06238703057169914, + 0.018582459539175034, + 0.05039922520518303, + 0.0050338744185864925, + 0.004406102001667023, + 0.009751887060701847, + 0.004808221943676472, + 0.005423519294708967, + 0.00479148468002677, + 0.0061711398884654045, + 0.005367688834667206, + 0.0025061212945729494, + 0.001873411238193512, + 0.0032204315066337585, + 0.005299681331962347, + 0.002423075493425131, + 0.002652163151651621, + 0.003721194341778755, + 0.002598253544420004, + 0.002931605326011777, + 0.0025851253885775805, + 0.003920271061360836, + 0.00370821263641119, + 0.004133640788495541, + 0.0027648548129945993, + 0.003879062132909894, + 0.0031172577291727066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.03371734917163849 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330356277525425 + ], + "opposite_logprob_tokens": [ + -3.773303508758545 + ], + "logprob_reward": -0.02330356277525425, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330356277525425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17772741615772247, + 0.06270460784435272, + 0.055456552654504776, + 0.08007901161909103, + 0.1629188507795334, + 0.12742404639720917, + 0.23824754357337952, + 0.029758449643850327, + 0.05283081904053688, + 0.012852786108851433 + ], + "credit_map": [ + 0.021450959146022797, + 0.024011554196476936, + 0.03689483925700188, + 0.013309592381119728, + 0.01255091279745102, + 0.013115372508764267, + 0.024546239525079727, + 0.03184795007109642, + 0.008910794742405415, + 0.006711879745125771, + 0.013558517210185528, + 0.006259534507989883, + 0.005819025915116072, + 0.004858388565480709, + 0.007159929722547531, + 0.009426540695130825, + 0.0060087330639362335, + 0.008027588948607445, + 0.01144329272210598, + 0.005570969078689814, + 0.00541211711242795, + 0.005655890330672264, + 0.0056831203401088715, + 0.007654838263988495, + 0.013488776050508022, + 0.007949349470436573, + 0.01743451878428459, + 0.008490181528031826, + 0.008042622357606888, + 0.0051200552843511105, + 0.009626783430576324, + 0.009926725178956985, + 0.015910295769572258, + 0.02305031567811966, + 0.04137122631072998, + 0.017745325341820717, + 0.01062435656785965, + 0.01778690703213215, + 0.019144902005791664, + 0.017285514622926712, + 0.014572760090231895, + 0.015115956775844097, + 0.019527211785316467, + 0.019614936783909798, + 0.011746818199753761, + 0.011657575145363808, + 0.014745264314115047, + 0.020443519577383995, + 0.008192240260541439, + 0.025425242260098457, + 0.01309769507497549, + 0.01602298580110073, + 0.0057396600022912025, + 0.07736755162477493, + 0.025653794407844543, + 0.06674838066101074, + 0.003968851175159216, + 0.0021073182579129934, + 0.00767773250117898, + 0.003303030738607049, + 0.0020451999735087156, + 0.003135033417493105, + 0.0031310657504945993, + 0.004390216898173094, + 0.008914824575185776, + 0.004764110781252384, + 0.015131821855902672, + 0.0045715379528701305, + 0.002832109108567238, + 0.004186131525784731, + 0.006141403689980507, + 0.0062888795509934425, + 0.0013633755734190345, + 0.0017512637423351407, + 0.0014870750019326806, + 0.001841315533965826, + 0.0013632270274683833, + 0.0016684150323271751, + 0.0017010952578857541, + 0.0016770182410255075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.02330356277525425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205760322511196 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.014205760322511196, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205760322511196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1724279224872589, + 0.03692092373967171, + 0.08371023833751678, + 0.0779452994465828, + 0.04920761659741402, + 0.06409397721290588, + 0.40350449085235596, + 0.03278771787881851, + 0.04339834302663803, + 0.03600350767374039 + ], + "credit_map": [ + 0.018142186105251312, + 0.021395524963736534, + 0.03346046432852745, + 0.024617191404104233, + 0.014820965938270092, + 0.020699216052889824, + 0.020581983029842377, + 0.01871037855744362, + 0.004874745383858681, + 0.004311275668442249, + 0.006788262166082859, + 0.004554980434477329, + 0.003774576587602496, + 0.0034751901403069496, + 0.004563768859952688, + 0.004578125663101673, + 0.011493518017232418, + 0.01227582711726427, + 0.012995446100831032, + 0.010531128384172916, + 0.00909815076738596, + 0.007419174071401358, + 0.010196973569691181, + 0.00970002543181181, + 0.008068347349762917, + 0.013407229445874691, + 0.01289516594260931, + 0.008795337751507759, + 0.009118303656578064, + 0.007818292826414108, + 0.010222038254141808, + 0.007620584219694138, + 0.0045954762026667595, + 0.007814045995473862, + 0.005933267530053854, + 0.009675032459199429, + 0.004539360757917166, + 0.005504818633198738, + 0.0056048971600830555, + 0.00554072018712759, + 0.006075384095311165, + 0.004848862066864967, + 0.014181041158735752, + 0.006780481897294521, + 0.01014514546841383, + 0.0057248142547905445, + 0.007857374846935272, + 0.008480872958898544, + 0.015087807551026344, + 0.044013459235429764, + 0.02320963703095913, + 0.029242364689707756, + 0.016553988680243492, + 0.12775349617004395, + 0.04159879684448242, + 0.1060449406504631, + 0.0053128208965063095, + 0.0018042526207864285, + 0.007393228355795145, + 0.003261141711845994, + 0.0028130272403359413, + 0.0034198127686977386, + 0.004079117439687252, + 0.004704318009316921, + 0.005660124588757753, + 0.003519307589158416, + 0.007025555241852999, + 0.008051211945712566, + 0.0039520845748484135, + 0.004686387721449137, + 0.00628917058929801, + 0.004214501474052668, + 0.0025085394736379385, + 0.003892329055815935, + 0.006181814707815647, + 0.006524580530822277, + 0.00475869607180357, + 0.0032962411642074585, + 0.0046802605502307415, + 0.004161048214882612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.014205760322511196 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium shot takes us to a deserted subway platform late at night. An elderly street performer takes center stage, wearing a well-worn, dark fedora, his graying hair tangled from beneath its brim. His face is lined with wrinkles as deep as a map. Clad in a heavy, dark gray wool coat with frayed edges, he closes his eyes in concentration, completely absorbed in the brass alto saxophone he holds. A few pale fluorescent lights on the ceiling provide the sole source of light, their cold, vertical beams illuminating his cheeks, which puff slightly from his intense playing, and casting a solitary, elongated shadow on the white tiled wall behind him. His fingers dance skillfully and briskly across the mother-of-pearl keys, warmed by constant use, his body swaying gently to the melancholic melody he produces. The melodious and slightly sad saxophone music echoes clearly in the empty underground space, interweaving with the faint sound of wind coming from the distant tunnel. The whole scene, set against the backdrop of cold colors, creates a quiet atmosphere of extreme loneliness yet full of artistic resilience.", + "seed": 10759, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_3/index_1465_seed_10759_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.035698776692152025, + "vlm_logprob": -0.035698776692152025, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.035698776692152025, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07596087455749512 + ], + "opposite_logprob_tokens": [ + -2.825960874557495 + ], + "logprob_reward": -0.07596087455749512, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07596087455749512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09023922681808472, + 0.05488865077495575, + 0.06496566534042358, + 0.04146414250135422, + 0.030130943283438683, + 0.025737261399626732, + 0.07782648503780365, + 0.14865292608737946, + 0.04837404936552048, + 0.41772061586380005 + ], + "credit_map": [ + 0.007559712044894695, + 0.01014463696628809, + 0.009747297503054142, + 0.009482283145189285, + 0.021548165008425713, + 0.011276712641119957, + 0.011139229871332645, + 0.009341186843812466, + 0.005567687097936869, + 0.007385289296507835, + 0.009299001656472683, + 0.005822261329740286, + 0.010960806161165237, + 0.005548274610191584, + 0.005054259207099676, + 0.005251069087535143, + 0.008433937095105648, + 0.0063741132616996765, + 0.012464657425880432, + 0.006667315028607845, + 0.007868830114603043, + 0.005580603610724211, + 0.009504063054919243, + 0.008072142489254475, + 0.0034224449191242456, + 0.003347408724948764, + 0.00923542957752943, + 0.007016647607088089, + 0.0038726881612092257, + 0.0034711784683167934, + 0.0067252484150230885, + 0.004373097326606512, + 0.0039622606709599495, + 0.003375488333404064, + 0.003698427462950349, + 0.0031606273259967566, + 0.004813280422240496, + 0.002936984645202756, + 0.004486979451030493, + 0.003696897067129612, + 0.0022912458516657352, + 0.003620862029492855, + 0.0024960986338555813, + 0.0021825097501277924, + 0.006339378654956818, + 0.0032547947484999895, + 0.002823192859068513, + 0.002729179570451379, + 0.011655613780021667, + 0.0073663629591465, + 0.016136111691594124, + 0.009441012516617775, + 0.007520740386098623, + 0.007853124290704727, + 0.008651593700051308, + 0.009201924316585064, + 0.010176628828048706, + 0.005437102168798447, + 0.012810475192964077, + 0.006735232193022966, + 0.005886044353246689, + 0.041249826550483704, + 0.017346974462270737, + 0.04901064932346344, + 0.004753204993903637, + 0.004199950490146875, + 0.010013100691139698, + 0.006281689740717411, + 0.004764799028635025, + 0.005332027096301317, + 0.006484828423708677, + 0.006544451229274273, + 0.03096175752580166, + 0.02140229381620884, + 0.035639241337776184, + 0.02394910529255867, + 0.013484585098922253, + 0.10331762582063675, + 0.0586274228990078, + 0.13033856451511383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.07596087455749512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370462030172348 + ], + "opposite_logprob_tokens": [ + -3.6513705253601074 + ], + "logprob_reward": -0.026370462030172348, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370462030172348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3032877743244171, + 0.12085480988025665, + 0.043912798166275024, + 0.020498216152191162, + 0.06096893176436424, + 0.025232577696442604, + 0.0272978488355875, + 0.061943765729665756, + 0.04750606790184975, + 0.2884971797466278 + ], + "credit_map": [ + 0.02910078503191471, + 0.027845706790685654, + 0.059344492852687836, + 0.0325079970061779, + 0.06291140615940094, + 0.02662554197013378, + 0.02901371195912361, + 0.035938143730163574, + 0.011973765678703785, + 0.017494602128863335, + 0.023100394755601883, + 0.013678179122507572, + 0.020399075001478195, + 0.010287771932780743, + 0.013304832391440868, + 0.010616197250783443, + 0.006232804153114557, + 0.005674303043633699, + 0.008182642050087452, + 0.0049665383994579315, + 0.0054509141482412815, + 0.0031105917878448963, + 0.005336239002645016, + 0.004958770237863064, + 0.0034670354798436165, + 0.0018682387890294194, + 0.003537722397595644, + 0.001906684716232121, + 0.0028713943902403116, + 0.0017979945987462997, + 0.0025583356618881226, + 0.002490811748430133, + 0.004982674494385719, + 0.005776315461844206, + 0.007992307655513287, + 0.006285140290856361, + 0.014492902904748917, + 0.006239810958504677, + 0.007743016351014376, + 0.00745677063241601, + 0.0025437581352889538, + 0.002800069050863385, + 0.0040063029155135155, + 0.0022187684662640095, + 0.005222757812589407, + 0.003035973059013486, + 0.0027775645721703768, + 0.002627384150400758, + 0.002530147787183523, + 0.00344130489975214, + 0.004772562999278307, + 0.0041387020610272884, + 0.004044197499752045, + 0.002561853500083089, + 0.003122872905805707, + 0.00268621020950377, + 0.005678573623299599, + 0.0059154159389436245, + 0.011279724538326263, + 0.007902756333351135, + 0.0068110087886452675, + 0.006084990222007036, + 0.007537257391959429, + 0.010734043084084988, + 0.0027150395326316357, + 0.0032222678419202566, + 0.006307696923613548, + 0.006294726859778166, + 0.008360289968550205, + 0.006850737147033215, + 0.007813907228410244, + 0.005941407755017281, + 0.01772029884159565, + 0.0069220708683133125, + 0.012538543902337551, + 0.009889126755297184, + 0.0114279855042696, + 0.08043680340051651, + 0.046272967010736465, + 0.103289395570755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.026370462030172348 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372138366103172 + ], + "opposite_logprob_tokens": [ + -3.408721446990967 + ], + "logprob_reward": -0.03372138366103172, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372138366103172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18823815882205963, + 0.1504155993461609, + 0.05370693281292915, + 0.03515688702464104, + 0.03520958498120308, + 0.025811659172177315, + 0.031569935381412506, + 0.22133509814739227, + 0.0766027420759201, + 0.18195338547229767 + ], + "credit_map": [ + 0.021053355187177658, + 0.03136969357728958, + 0.029908861964941025, + 0.028728898614645004, + 0.022501425817608833, + 0.019792119041085243, + 0.017960725352168083, + 0.016923075541853905, + 0.01439574547111988, + 0.02095438912510872, + 0.02944209985435009, + 0.01635737158358097, + 0.026239411905407906, + 0.016340816393494606, + 0.014516769908368587, + 0.012168999761343002, + 0.004686645232141018, + 0.0039507183246314526, + 0.013318490236997604, + 0.005208497866988182, + 0.01030655112117529, + 0.005486839450895786, + 0.006203594617545605, + 0.004545601084828377, + 0.002645589876919985, + 0.004962126724421978, + 0.004803935997188091, + 0.004395464435219765, + 0.004922001156955957, + 0.0032462971284985542, + 0.0054492102935910225, + 0.0047322604805231094, + 0.0025810357183218002, + 0.005716879386454821, + 0.004364163614809513, + 0.005206814967095852, + 0.004992310423403978, + 0.003166914451867342, + 0.004545053467154503, + 0.00463641295209527, + 0.003553144633769989, + 0.002598918741568923, + 0.0039017812814563513, + 0.002302480861544609, + 0.0043935757130384445, + 0.003045523539185524, + 0.0031920140609145164, + 0.002824219409376383, + 0.005303232930600643, + 0.0034277450758963823, + 0.006489915773272514, + 0.0030498611740767956, + 0.003821226768195629, + 0.002974731381982565, + 0.0032074882183223963, + 0.0032957354560494423, + 0.014395132660865784, + 0.005063314456492662, + 0.008962582796812057, + 0.010545894503593445, + 0.008633953519165516, + 0.06604879349470139, + 0.028310371562838554, + 0.07937506586313248, + 0.005848990753293037, + 0.005880274344235659, + 0.007856031879782677, + 0.013126387260854244, + 0.01690845936536789, + 0.006704930681735277, + 0.012452577240765095, + 0.007825089618563652, + 0.010971934534609318, + 0.005161788780242205, + 0.008922071196138859, + 0.009163379669189453, + 0.007764216512441635, + 0.04872416704893112, + 0.028296969830989838, + 0.06294884532690048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.03372138366103172 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636036090552807 + ], + "opposite_logprob_tokens": [ + -3.651360273361206 + ], + "logprob_reward": -0.02636036090552807, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636036090552807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2855244278907776, + 0.038621872663497925, + 0.06371026486158371, + 0.044249292463064194, + 0.031114721670746803, + 0.024274276569485664, + 0.04549386352300644, + 0.16186246275901794, + 0.07444624602794647, + 0.23070253431797028 + ], + "credit_map": [ + 0.028455359861254692, + 0.036959968507289886, + 0.03969433531165123, + 0.02688339352607727, + 0.07349091023206711, + 0.027669664472341537, + 0.0265530813485384, + 0.025817718356847763, + 0.0033032994251698256, + 0.0039151632227003574, + 0.009287767112255096, + 0.005743591580539942, + 0.003219547914341092, + 0.004488796461373568, + 0.004332848824560642, + 0.004330858122557402, + 0.007014908827841282, + 0.009888584725558758, + 0.011691314168274403, + 0.007774391211569309, + 0.008979215286672115, + 0.004704383201897144, + 0.0065390560775995255, + 0.007118411362171173, + 0.003724966896697879, + 0.003508772002533078, + 0.00835399143397808, + 0.007467284798622131, + 0.0056998636573553085, + 0.004293856676667929, + 0.0065208435989916325, + 0.004679716657847166, + 0.0025046567898243666, + 0.004013219848275185, + 0.003890648251399398, + 0.0036198580637574196, + 0.007342095021158457, + 0.0038605888839811087, + 0.003345509059727192, + 0.0025381490122526884, + 0.0017411602893844247, + 0.002378851408138871, + 0.0037109442055225372, + 0.0028562555089592934, + 0.0032416072208434343, + 0.002488119760528207, + 0.0036031834315508604, + 0.004254153463989496, + 0.0060841795057058334, + 0.0036338353529572487, + 0.01133584976196289, + 0.0049559930339455605, + 0.00514490669593215, + 0.0048391567543148994, + 0.004744338802993298, + 0.0047556012868881226, + 0.01760607212781906, + 0.006080857943743467, + 0.01550705824047327, + 0.014784331433475018, + 0.008584879338741302, + 0.0338016115128994, + 0.0187231358140707, + 0.04677450656890869, + 0.013996965251863003, + 0.006257521919906139, + 0.013022113591432571, + 0.006694056559354067, + 0.005168837495148182, + 0.01153600588440895, + 0.009124967269599438, + 0.008645775727927685, + 0.021531276404857635, + 0.013517647050321102, + 0.025082029402256012, + 0.013016453012824059, + 0.009413449093699455, + 0.054869815707206726, + 0.029428068548440933, + 0.06384380161762238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.02636036090552807 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608080230653286 + ], + "opposite_logprob_tokens": [ + -4.141080856323242 + ], + "logprob_reward": -0.01608080230653286, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608080230653286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13831554353237152, + 0.12113280594348907, + 0.031409360468387604, + 0.04022226482629776, + 0.0964885875582695, + 0.037963639944791794, + 0.06871769577264786, + 0.20763716101646423, + 0.055572617799043655, + 0.2025403380393982 + ], + "credit_map": [ + 0.020058119669556618, + 0.0187604408711195, + 0.0174957737326622, + 0.014266709797084332, + 0.025121649727225304, + 0.01482534408569336, + 0.01571543700993061, + 0.012072071433067322, + 0.01186220720410347, + 0.014994893223047256, + 0.024810494855046272, + 0.016707487404346466, + 0.019952448084950447, + 0.010267608799040318, + 0.012460584752261639, + 0.010077082552015781, + 0.001842002384364605, + 0.004076131619513035, + 0.003498712321743369, + 0.006044996436685324, + 0.005384011659771204, + 0.0027875136584043503, + 0.004607329145073891, + 0.003168660681694746, + 0.002909418661147356, + 0.00788784958422184, + 0.003771141404286027, + 0.005048528779298067, + 0.007630530744791031, + 0.004126276355236769, + 0.004928607027977705, + 0.003919912036508322, + 0.010224110446870327, + 0.017134856432676315, + 0.013263692148029804, + 0.011283580213785172, + 0.014140831306576729, + 0.007668071426451206, + 0.012261922471225262, + 0.01051152404397726, + 0.004609930794686079, + 0.004314107354730368, + 0.004935322795063257, + 0.003210514085367322, + 0.006452007684856653, + 0.005015015602111816, + 0.0045977067202329636, + 0.004829037934541702, + 0.008896199986338615, + 0.005511966068297625, + 0.016812337562441826, + 0.006248614750802517, + 0.010193203575909138, + 0.007841629907488823, + 0.007039903197437525, + 0.006173838395625353, + 0.01410695817321539, + 0.004088656511157751, + 0.009513601660728455, + 0.006063761655241251, + 0.007364743389189243, + 0.05928005278110504, + 0.027821404859423637, + 0.07939799129962921, + 0.006214587949216366, + 0.0038083309773355722, + 0.009290331043303013, + 0.008195039816200733, + 0.010359995998442173, + 0.004226273391395807, + 0.007373330648988485, + 0.006104730069637299, + 0.012998529709875584, + 0.005117947701364756, + 0.008136137388646603, + 0.009331230074167252, + 0.01143556460738182, + 0.05379130318760872, + 0.03158675879240036, + 0.07014285027980804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.01608080230653286 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium shot takes us to a deserted subway platform late at night. An elderly street performer takes center stage, wearing a well-worn, dark fedora, his graying hair tangled from beneath its brim. His face is lined with wrinkles as deep as a map. Clad in a heavy, dark gray wool coat with frayed edges, he closes his eyes in concentration, completely absorbed in the brass alto saxophone he holds. A few pale fluorescent lights on the ceiling provide the sole source of light, their cold, vertical beams illuminating his cheeks, which puff slightly from his intense playing, and casting a solitary, elongated shadow on the white tiled wall behind him. His fingers dance skillfully and briskly across the mother-of-pearl keys, warmed by constant use, his body swaying gently to the melancholic melody he produces. The melodious and slightly sad saxophone music echoes clearly in the empty underground space, interweaving with the faint sound of wind coming from the distant tunnel. The whole scene, set against the backdrop of cold colors, creates a quiet atmosphere of extreme loneliness yet full of artistic resilience.", + "seed": 10760, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_3/index_1465_seed_10760_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03008001670241356, + "vlm_logprob": -0.03008001670241356, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03008001670241356, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08474630117416382 + ], + "opposite_logprob_tokens": [ + -2.8347463607788086 + ], + "logprob_reward": -0.08474630117416382, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08474630117416382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16114503145217896, + 0.045276038348674774, + 0.13027261197566986, + 0.11495644599199295, + 0.0979224443435669, + 0.03748730570077896, + 0.06385325640439987, + 0.18221323192119598, + 0.0696258693933487, + 0.09724783152341843 + ], + "credit_map": [ + 0.01048175897449255, + 0.012173570692539215, + 0.017580512911081314, + 0.03952503949403763, + 0.013849333859980106, + 0.016002001240849495, + 0.035461898893117905, + 0.016070902347564697, + 0.0035478679928928614, + 0.004215487744659185, + 0.005268570967018604, + 0.005625415593385696, + 0.006498691160231829, + 0.006832776125520468, + 0.004969063214957714, + 0.008318166248500347, + 0.008906437084078789, + 0.017331132665276527, + 0.015777695924043655, + 0.015230352990329266, + 0.015715621411800385, + 0.02048158273100853, + 0.019035566598176956, + 0.017794227227568626, + 0.004610180854797363, + 0.005092450883239508, + 0.006032597739249468, + 0.02696930803358555, + 0.005855998955667019, + 0.013358457945287228, + 0.010527356527745724, + 0.042510103434324265, + 0.004849245771765709, + 0.015715522691607475, + 0.013728140853345394, + 0.018553294241428375, + 0.00855614710599184, + 0.009789944626390934, + 0.01642192341387272, + 0.010308222845196724, + 0.002645139815285802, + 0.006196216680109501, + 0.00355465873144567, + 0.004820731934159994, + 0.005721039138734341, + 0.004787574987858534, + 0.0059683253057301044, + 0.003793619340285659, + 0.004384721163660288, + 0.008347908966243267, + 0.010457547381520271, + 0.010511868633329868, + 0.0071317097172141075, + 0.006002729758620262, + 0.009980804286897182, + 0.007035972084850073, + 0.009568646550178528, + 0.014799943193793297, + 0.025298375636339188, + 0.027209589257836342, + 0.021514756605029106, + 0.030857441946864128, + 0.02189193107187748, + 0.03107253648340702, + 0.006354291457682848, + 0.008294248953461647, + 0.012143907137215137, + 0.010690138675272465, + 0.009431046433746815, + 0.006904651410877705, + 0.008980748243629932, + 0.006826835684478283, + 0.004676387645304203, + 0.010990605689585209, + 0.014420127496123314, + 0.012360112741589546, + 0.01719418354332447, + 0.015335245057940483, + 0.008068447932600975, + 0.014202717691659927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.08474630117416382 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821083389222622 + ], + "opposite_logprob_tokens": [ + -4.0182108879089355 + ], + "logprob_reward": -0.01821083389222622, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821083389222622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16820834577083588, + 0.1505083292722702, + 0.05391322076320648, + 0.30209851264953613, + 0.09047587215900421, + 0.019311219453811646, + 0.05245741456747055, + 0.07671063393354416, + 0.03303659334778786, + 0.053279850631952286 + ], + "credit_map": [ + 0.014238322153687477, + 0.012680086307227612, + 0.01488206535577774, + 0.037742964923381805, + 0.018830372020602226, + 0.013153383508324623, + 0.037935081869363785, + 0.018746059387922287, + 0.010465013794600964, + 0.016144024208188057, + 0.016021279618144035, + 0.019200533628463745, + 0.02034691348671913, + 0.030782848596572876, + 0.018587887287139893, + 0.018959829583764076, + 0.0038682755548506975, + 0.005950824357569218, + 0.006696128286421299, + 0.007611551322042942, + 0.0070025245659053326, + 0.005611063446849585, + 0.00877956859767437, + 0.008393283933401108, + 0.011843940243124962, + 0.01641024649143219, + 0.0191735178232193, + 0.0665811151266098, + 0.013230454176664352, + 0.034035470336675644, + 0.03245329484343529, + 0.10837048292160034, + 0.004415364935994148, + 0.010273955762386322, + 0.01743004098534584, + 0.013187773525714874, + 0.008904273621737957, + 0.010787148028612137, + 0.01270108949393034, + 0.012776223942637444, + 0.001305407378822565, + 0.0033859757240861654, + 0.0021057988051325083, + 0.0025656376965343952, + 0.0026129113975912333, + 0.0015277444617822766, + 0.0030134490225464106, + 0.0027942946180701256, + 0.0017638722201809287, + 0.004016051068902016, + 0.006328550633043051, + 0.007910311222076416, + 0.008029437623918056, + 0.003745641093701124, + 0.013833736069500446, + 0.006829817313700914, + 0.004426516592502594, + 0.0036663629580289125, + 0.006232719402760267, + 0.013730190694332123, + 0.007825764827430248, + 0.013826592825353146, + 0.009896166622638702, + 0.017106318846344948, + 0.0016300524584949017, + 0.003172023454681039, + 0.0044535077176988125, + 0.004802754148840904, + 0.0041734082624316216, + 0.0031833588145673275, + 0.007510067895054817, + 0.004111419431865215, + 0.0023029756266623735, + 0.008868067525327206, + 0.006391655653715134, + 0.00702247628942132, + 0.009572509676218033, + 0.007027451414614916, + 0.005634999368339777, + 0.006459713447839022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.01821083389222622 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596321672201157 + ], + "opposite_logprob_tokens": [ + -3.8955962657928467 + ], + "logprob_reward": -0.020596321672201157, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596321672201157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1915840357542038, + 0.039676763117313385, + 0.07033060491085052, + 0.49369317293167114, + 0.01820414513349533, + 0.018568729981780052, + 0.023574667051434517, + 0.04099566861987114, + 0.07726659625768661, + 0.026105590164661407 + ], + "credit_map": [ + 0.01286184974014759, + 0.013688082806766033, + 0.01615043170750141, + 0.042693935334682465, + 0.024105383083224297, + 0.017517371103167534, + 0.04447776824235916, + 0.020089227706193924, + 0.003662463743239641, + 0.004654914606362581, + 0.005329719744622707, + 0.004779935348778963, + 0.0045735882595181465, + 0.005462820176035166, + 0.0051786149851977825, + 0.006034705322235823, + 0.005232009571045637, + 0.010138475336134434, + 0.006965182721614838, + 0.008980854414403439, + 0.009957472793757915, + 0.012288344092667103, + 0.009510396979749203, + 0.007257866207510233, + 0.021295109763741493, + 0.023300854489207268, + 0.02885730005800724, + 0.1161528006196022, + 0.020639585331082344, + 0.06021500378847122, + 0.044252436608076096, + 0.17898006737232208, + 0.0010197758674621582, + 0.0024163888301700354, + 0.0021592353004962206, + 0.0024181604385375977, + 0.0021697066258639097, + 0.0018195930169895291, + 0.002509679412469268, + 0.003691605059430003, + 0.0016614077612757683, + 0.0026622777804732323, + 0.0024331267923116684, + 0.002635242650285363, + 0.0019543457310646772, + 0.0020342536736279726, + 0.002498633461073041, + 0.0026894421316683292, + 0.0013255688827484846, + 0.003187012393027544, + 0.0027162085752934217, + 0.0030989667866379023, + 0.004571661818772554, + 0.002165627432987094, + 0.004328873474150896, + 0.00218074768781662, + 0.0023393668234348297, + 0.003962129354476929, + 0.005045893602073193, + 0.005780515726655722, + 0.005930987652391195, + 0.006102250888943672, + 0.005581056233495474, + 0.006253467407077551, + 0.003518515732139349, + 0.010074629448354244, + 0.007390521466732025, + 0.010565580800175667, + 0.019211996346712112, + 0.006423605605959892, + 0.013622774742543697, + 0.006458982825279236, + 0.0011639549629762769, + 0.004260006360709667, + 0.002957233227789402, + 0.0037404370959848166, + 0.004534728825092316, + 0.0034959022887051105, + 0.003126344410702586, + 0.0028269838076084852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.020596321672201157 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198776990175247 + ], + "opposite_logprob_tokens": [ + -4.0181989669799805 + ], + "logprob_reward": -0.018198776990175247, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018198776990175247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08562169224023819, + 0.027917804196476936, + 0.0819210410118103, + 0.3371759355068207, + 0.08165914565324783, + 0.017678869888186455, + 0.055818382650613785, + 0.04862343147397041, + 0.12982632219791412, + 0.13375739753246307 + ], + "credit_map": [ + 0.007369005586951971, + 0.006092933472245932, + 0.007277307566255331, + 0.021909592673182487, + 0.0088088009506464, + 0.004595409147441387, + 0.02270323596894741, + 0.006865398027002811, + 0.00228484277613461, + 0.003142139408737421, + 0.0035703645553439856, + 0.003668047022074461, + 0.003667994635179639, + 0.004796040244400501, + 0.0037600912619382143, + 0.0030282812658697367, + 0.005102342460304499, + 0.012261652387678623, + 0.010427896864712238, + 0.01107732579112053, + 0.01247753668576479, + 0.008742254227399826, + 0.011958003044128418, + 0.009874020703136921, + 0.013509727083146572, + 0.018391365185379982, + 0.02054074965417385, + 0.07529205828905106, + 0.015174901112914085, + 0.038197729736566544, + 0.03551766276359558, + 0.12055172771215439, + 0.004047094378620386, + 0.01217099092900753, + 0.014275179244577885, + 0.009489067830145359, + 0.008632107637822628, + 0.006282430607825518, + 0.011430121958255768, + 0.01533213909715414, + 0.0012716143392026424, + 0.0031809669453650713, + 0.0020406439434736967, + 0.001750962808728218, + 0.002444973448291421, + 0.002145790494978428, + 0.0027327791322022676, + 0.0021111364476382732, + 0.0050405352376401424, + 0.004739219322800636, + 0.009750111028552055, + 0.007147803902626038, + 0.007056558970361948, + 0.00885437335819006, + 0.006989572662860155, + 0.006240201182663441, + 0.0052604698576033115, + 0.0039270613342523575, + 0.005990886129438877, + 0.0060761249624192715, + 0.006122288294136524, + 0.00817240122705698, + 0.006308446638286114, + 0.006765751633793116, + 0.007437833119183779, + 0.01432965137064457, + 0.0213365126401186, + 0.018551887944340706, + 0.015571783296763897, + 0.01357257179915905, + 0.02383067086338997, + 0.015195406973361969, + 0.005096124019473791, + 0.01891098916530609, + 0.017289385199546814, + 0.0171516053378582, + 0.030677447095513344, + 0.01837063394486904, + 0.011816089041531086, + 0.014445101842284203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.018198776990175247 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647849783301353 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008647849783301353, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647849783301353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12776900827884674, + 0.027036607265472412, + 0.024081572890281677, + 0.580348789691925, + 0.035457391291856766, + 0.020663263276219368, + 0.02099478803575039, + 0.023161379620432854, + 0.08931203186511993, + 0.05117509886622429 + ], + "credit_map": [ + 0.0080028111115098, + 0.007866896688938141, + 0.009479670785367489, + 0.03308973088860512, + 0.010262181051075459, + 0.010079258121550083, + 0.03550034388899803, + 0.013488112017512321, + 0.0014721084153279662, + 0.002084102714434266, + 0.004382945131510496, + 0.004079781007021666, + 0.003938230685889721, + 0.0028171504382044077, + 0.004738843068480492, + 0.0035234459210187197, + 0.0015572232659906149, + 0.002311922609806061, + 0.0036566276103258133, + 0.0035063603427261114, + 0.004160101059824228, + 0.0021561093162745237, + 0.0034195894841104746, + 0.0033136396668851376, + 0.022041380405426025, + 0.026714632287621498, + 0.030805394053459167, + 0.13608025014400482, + 0.01942780800163746, + 0.06872136145830154, + 0.05576179921627045, + 0.22079619765281677, + 0.0018043417949229479, + 0.004998629447072744, + 0.004404227714985609, + 0.006235211156308651, + 0.003910311032086611, + 0.003470122581347823, + 0.005518399644643068, + 0.005116146989166737, + 0.0014749816618859768, + 0.002850969322025776, + 0.0020953950006514788, + 0.003251645714044571, + 0.0029067392461001873, + 0.00210963049903512, + 0.003251961898058653, + 0.0027219406329095364, + 0.0019462978234514594, + 0.002044643508270383, + 0.0029684395994991064, + 0.0025760564021766186, + 0.0026779037434607744, + 0.0033654002472758293, + 0.003278705757111311, + 0.0021373413037508726, + 0.001428382471203804, + 0.0028496708255261183, + 0.00346676679328084, + 0.003634158056229353, + 0.0026038605719804764, + 0.0033136054407805204, + 0.003033645451068878, + 0.0028312895447015762, + 0.004266977775841951, + 0.011508669704198837, + 0.009763199836015701, + 0.012210349552333355, + 0.02065436728298664, + 0.00804112758487463, + 0.014354173094034195, + 0.00851316936314106, + 0.0020952336490154266, + 0.00834767334163189, + 0.00674305809661746, + 0.007039931137114763, + 0.010731302201747894, + 0.006067306734621525, + 0.00470169959589839, + 0.005448896437883377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.008647849783301353 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium shot takes us to a deserted subway platform late at night. An elderly street performer takes center stage, wearing a well-worn, dark fedora, his graying hair tangled from beneath its brim. His face is lined with wrinkles as deep as a map. Clad in a heavy, dark gray wool coat with frayed edges, he closes his eyes in concentration, completely absorbed in the brass alto saxophone he holds. A few pale fluorescent lights on the ceiling provide the sole source of light, their cold, vertical beams illuminating his cheeks, which puff slightly from his intense playing, and casting a solitary, elongated shadow on the white tiled wall behind him. His fingers dance skillfully and briskly across the mother-of-pearl keys, warmed by constant use, his body swaying gently to the melancholic melody he produces. The melodious and slightly sad saxophone music echoes clearly in the empty underground space, interweaving with the faint sound of wind coming from the distant tunnel. The whole scene, set against the backdrop of cold colors, creates a quiet atmosphere of extreme loneliness yet full of artistic resilience.", + "seed": 10761, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_3/index_1465_seed_10761_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.038782715797424316, + "vlm_logprob": -0.038782715797424316, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.038782715797424316, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08566499501466751 + ], + "opposite_logprob_tokens": [ + -2.710664987564087 + ], + "logprob_reward": -0.08566499501466751, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08566499501466751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26347920298576355, + 0.061230678111314774, + 0.07604032754898071, + 0.0984659492969513, + 0.1461685448884964, + 0.03530709445476532, + 0.22046487033367157, + 0.02908368408679962, + 0.02377622202038765, + 0.04598339647054672 + ], + "credit_map": [ + 0.02527184784412384, + 0.015199691988527775, + 0.03871261700987816, + 0.031894296407699585, + 0.03570587560534477, + 0.03857475519180298, + 0.038430340588092804, + 0.03968975692987442, + 0.00535240862518549, + 0.0031992411240935326, + 0.010163690894842148, + 0.00817121472209692, + 0.011053349822759628, + 0.005094187334179878, + 0.011066759005188942, + 0.0071298242546617985, + 0.006548842415213585, + 0.006105483043938875, + 0.014957917854189873, + 0.009683388285338879, + 0.009763434529304504, + 0.008863774128258228, + 0.010232671163976192, + 0.009884815663099289, + 0.007992328144609928, + 0.006940500345081091, + 0.014607438817620277, + 0.016527647152543068, + 0.015722626820206642, + 0.010513894259929657, + 0.012555037625133991, + 0.013606477528810501, + 0.01751350425183773, + 0.02074340730905533, + 0.021356001496315002, + 0.021728046238422394, + 0.017837373539805412, + 0.012211129069328308, + 0.019449539482593536, + 0.015329539775848389, + 0.003983020316809416, + 0.0041388473473489285, + 0.005909808445721865, + 0.005074824206531048, + 0.004463233519345522, + 0.002899660263210535, + 0.0050524636171758175, + 0.0037852381356060505, + 0.015852676704525948, + 0.023955421522259712, + 0.04356778413057327, + 0.029757963493466377, + 0.031177325174212456, + 0.024967215955257416, + 0.02573201060295105, + 0.02545447275042534, + 0.002387142274528742, + 0.0013833573320880532, + 0.0033857899252325296, + 0.005757720209658146, + 0.0038359740283340216, + 0.0026788278482854366, + 0.005928890313953161, + 0.0037259829696267843, + 0.0019965963438153267, + 0.0012886931654065847, + 0.0019985211547464132, + 0.00326559622772038, + 0.0045065428130328655, + 0.00200367602519691, + 0.00567262526601553, + 0.003043971722945571, + 0.004677098244428635, + 0.0033417565282434225, + 0.007073270622640848, + 0.006600555498152971, + 0.006091927643865347, + 0.0042107366025447845, + 0.007703565061092377, + 0.006284484639763832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.08566499501466751 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026371970772743225 + ], + "opposite_logprob_tokens": [ + -3.651371955871582 + ], + "logprob_reward": -0.026371970772743225, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026371970772743225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3537314236164093, + 0.12468871474266052, + 0.15056222677230835, + 0.050311096012592316, + 0.05195511505007744, + 0.05594484135508537, + 0.08100616186857224, + 0.03380929306149483, + 0.038629159331321716, + 0.059362005442380905 + ], + "credit_map": [ + 0.03584999591112137, + 0.015624761581420898, + 0.0442754402756691, + 0.049663424491882324, + 0.05153166502714157, + 0.046917494386434555, + 0.06763215363025665, + 0.04223645478487015, + 0.009749816730618477, + 0.006499753333628178, + 0.011898079887032509, + 0.013739174231886864, + 0.03055848740041256, + 0.011814212426543236, + 0.02512209862470627, + 0.015307078137993813, + 0.008642969653010368, + 0.006661119870841503, + 0.012115208432078362, + 0.018631821498274803, + 0.030636005103588104, + 0.01228097453713417, + 0.03487258777022362, + 0.026721514761447906, + 0.0037800457794219255, + 0.004075305070728064, + 0.009298249147832394, + 0.008303851820528507, + 0.007128055673092604, + 0.005504020489752293, + 0.007491826545447111, + 0.004729738924652338, + 0.006120078731328249, + 0.003766163717955351, + 0.006860812194645405, + 0.008332153782248497, + 0.008680307306349277, + 0.004619044717401266, + 0.00819058995693922, + 0.005385960917919874, + 0.005284572951495647, + 0.00571438716724515, + 0.008868162520229816, + 0.005613114219158888, + 0.007403592579066753, + 0.00802785437554121, + 0.00816104095429182, + 0.006872115656733513, + 0.005196386482566595, + 0.011145221069455147, + 0.01160747092217207, + 0.01148935966193676, + 0.011404811404645443, + 0.009740585461258888, + 0.009841576218605042, + 0.010580741800367832, + 0.0027176577132195234, + 0.0020770193077623844, + 0.004322899039834738, + 0.004771007224917412, + 0.005430289078503847, + 0.002984815277159214, + 0.005033747758716345, + 0.006471855100244284, + 0.003087433986365795, + 0.0017653141403570771, + 0.003482201835140586, + 0.006093637552112341, + 0.005962558090686798, + 0.003391280770301819, + 0.006043357774615288, + 0.008803372271358967, + 0.0035869553685188293, + 0.0026685709599405527, + 0.005439471919089556, + 0.0132218599319458, + 0.012374051846563816, + 0.004006001632660627, + 0.010121555998921394, + 0.0079435333609581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.026371970772743225 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308338090777397 + ], + "opposite_logprob_tokens": [ + -3.168083429336548 + ], + "logprob_reward": -0.04308338090777397, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308338090777397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15550732612609863, + 0.10376378893852234, + 0.062135688960552216, + 0.06413762271404266, + 0.07449135929346085, + 0.12653009593486786, + 0.22584658861160278, + 0.07131753861904144, + 0.03869149461388588, + 0.07757849991321564 + ], + "credit_map": [ + 0.015459536574780941, + 0.008548476733267307, + 0.014700083993375301, + 0.024888649582862854, + 0.023500295355916023, + 0.012045522220432758, + 0.02856721170246601, + 0.027797549962997437, + 0.006058497354388237, + 0.004437949042767286, + 0.0092965979129076, + 0.010927611961960793, + 0.022414008155465126, + 0.006544553674757481, + 0.03185505047440529, + 0.012229502201080322, + 0.003684352384880185, + 0.00408061733469367, + 0.010939734056591988, + 0.008725481107831001, + 0.011827229522168636, + 0.006181308999657631, + 0.009657619521021843, + 0.007039344869554043, + 0.005165738984942436, + 0.0050141466781497, + 0.007773907855153084, + 0.008589130826294422, + 0.010945171117782593, + 0.009259000420570374, + 0.008262746967375278, + 0.009127778001129627, + 0.006261597853153944, + 0.007345316931605339, + 0.011971176601946354, + 0.0088777681812644, + 0.013532210141420364, + 0.008384576067328453, + 0.009842737577855587, + 0.008275972679257393, + 0.011336023919284344, + 0.014640112407505512, + 0.017162073403596878, + 0.014148425310850143, + 0.021448079496622086, + 0.013399329036474228, + 0.018132731318473816, + 0.016263317316770554, + 0.01640179567039013, + 0.028824971988797188, + 0.0303211472928524, + 0.028598159551620483, + 0.03799734264612198, + 0.02363390475511551, + 0.028386717662215233, + 0.03168253228068352, + 0.00668220454826951, + 0.004247706849128008, + 0.008495617657899857, + 0.008290057070553303, + 0.01460321806371212, + 0.007026746869087219, + 0.009252963587641716, + 0.012719019316136837, + 0.004149856045842171, + 0.0020084725692868233, + 0.0027080823201686144, + 0.00429915776476264, + 0.006530901417136192, + 0.0030706552788615227, + 0.007479846011847258, + 0.00844451691955328, + 0.005138770677149296, + 0.003264412982389331, + 0.006184575613588095, + 0.018606916069984436, + 0.010530350729823112, + 0.004968446679413319, + 0.018561920151114464, + 0.010323097929358482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.04308338090777397 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020593050867319107 + ], + "opposite_logprob_tokens": [ + -3.8955931663513184 + ], + "logprob_reward": -0.020593050867319107, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020593050867319107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13734370470046997, + 0.08626656234264374, + 0.027270840480923653, + 0.10185302048921585, + 0.33695855736732483, + 0.09298735111951828, + 0.044738106429576874, + 0.03226302191615105, + 0.0365055613219738, + 0.10381326079368591 + ], + "credit_map": [ + 0.013822397217154503, + 0.010233264416456223, + 0.02063484862446785, + 0.020198801532387733, + 0.014314700849354267, + 0.02164458855986595, + 0.01753576286137104, + 0.01895933970808983, + 0.0075537157244980335, + 0.004857890773564577, + 0.010173380374908447, + 0.009947232902050018, + 0.014025245793163776, + 0.01186191663146019, + 0.016191991046071053, + 0.011655188165605068, + 0.002325415378436446, + 0.002115840557962656, + 0.003747499780729413, + 0.00340096652507782, + 0.004688340704888105, + 0.0022271962370723486, + 0.004484946373850107, + 0.0042806351557374, + 0.008561990223824978, + 0.007582223042845726, + 0.016372034326195717, + 0.015743782743811607, + 0.015617186203598976, + 0.012550419196486473, + 0.013777499087154865, + 0.011647884733974934, + 0.03912115842103958, + 0.038914501667022705, + 0.06838655471801758, + 0.04342961683869362, + 0.033390339463949203, + 0.039233189076185226, + 0.04613017290830612, + 0.028353018686175346, + 0.009907322935760021, + 0.011206802912056446, + 0.017527252435684204, + 0.011611858382821083, + 0.009656662121415138, + 0.011672467924654484, + 0.012965659610927105, + 0.008439334109425545, + 0.00479410495609045, + 0.004345405381172895, + 0.006573505699634552, + 0.00650437455624342, + 0.004367387387901545, + 0.004050042480230331, + 0.006853150669485331, + 0.007250132039189339, + 0.0032773930579423904, + 0.002006669994443655, + 0.0059636058285832405, + 0.00488076638430357, + 0.004411958158016205, + 0.003194078803062439, + 0.003509246278554201, + 0.005019302945584059, + 0.0027218079194426537, + 0.0018232879228889942, + 0.003702946240082383, + 0.0031170265283435583, + 0.006282418966293335, + 0.0026396934408694506, + 0.0071456776931881905, + 0.009072700515389442, + 0.012695273384451866, + 0.010487854480743408, + 0.018542824313044548, + 0.013322580605745316, + 0.010576039552688599, + 0.010541950352489948, + 0.015127155929803848, + 0.012519575655460358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.020593050867319107 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted subway platform late at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure an elderly man wearing a dark fedora and a dark gray wool coat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man playing a brass alto saxophone with his eyes closed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the only light sources pale fluorescent ceiling lights casting vertical beams and an elongated shadow on a white tiled wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the man's body swaying gently while his fingers move briskly across the saxophone keys?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200181424617767 + ], + "opposite_logprob_tokens": [ + -4.018200397491455 + ], + "logprob_reward": -0.018200181424617767, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018200181424617767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1522352695465088, + 0.0934973657131195, + 0.04025335609912872, + 0.05883367359638214, + 0.12336848676204681, + 0.17491094768047333, + 0.12368905544281006, + 0.07190673053264618, + 0.10073781758546829, + 0.06056733429431915 + ], + "credit_map": [ + 0.01626822166144848, + 0.009140490554273129, + 0.016546543687582016, + 0.022071829065680504, + 0.024983854964375496, + 0.011770923621952534, + 0.02151944860816002, + 0.029933959245681763, + 0.007020927965641022, + 0.0031175778713077307, + 0.013198046945035458, + 0.017590133473277092, + 0.018991325050592422, + 0.007656535599380732, + 0.016160624101758003, + 0.009762195870280266, + 0.0032685163896530867, + 0.0018469244241714478, + 0.005973226856440306, + 0.005317932926118374, + 0.008305320516228676, + 0.004009181167930365, + 0.00728497514501214, + 0.00424727750942111, + 0.004361191764473915, + 0.0029791516717523336, + 0.005616809707134962, + 0.012016172520816326, + 0.012714422307908535, + 0.005762432236224413, + 0.009164344519376755, + 0.0062191481702029705, + 0.014422418549656868, + 0.015049775131046772, + 0.0215372946113348, + 0.018337612971663475, + 0.015345747582614422, + 0.009627733379602432, + 0.01655910536646843, + 0.012488805688917637, + 0.016051942482590675, + 0.014185994863510132, + 0.0381472148001194, + 0.02214961312711239, + 0.015386824496090412, + 0.021996762603521347, + 0.02773246169090271, + 0.019260121509432793, + 0.009477975778281689, + 0.015147527679800987, + 0.013748195953667164, + 0.0173017755150795, + 0.019978243857622147, + 0.012185651808977127, + 0.01835721544921398, + 0.017492465674877167, + 0.00833787489682436, + 0.004623423330485821, + 0.011413374915719032, + 0.007630550768226385, + 0.014868898317217827, + 0.004838837310671806, + 0.011453220620751381, + 0.008740551769733429, + 0.007402428891509771, + 0.003107805037871003, + 0.005631742067635059, + 0.01333841122686863, + 0.018916534259915352, + 0.006448320113122463, + 0.02465820498764515, + 0.021234367042779922, + 0.005957570858299732, + 0.0027048152405768633, + 0.007242036517709494, + 0.009670967236161232, + 0.009918959811329842, + 0.00502479262650013, + 0.013193684630095959, + 0.006854505278170109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.018200181424617767 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_4/index_tensor([5832, 5832])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_4/index_tensor([5832, 5832])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..51dd977eb9b2ea0e56973c63aaabfc753d00013a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_4/index_tensor([5832, 5832])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a panoramic high-angle drone perspective, the camera slowly and steadily glides right, tracing the path of a lone ski run carved into a vast, pristine mountainside covered in deep, untouched powder snow. The track itself is a delicate serpentine line, a graceful S-curve along the steep slope, its two parallel grooves sharply defined against the otherwise pristine white canvas. The low-angle rays of the late afternoon sun illuminate the scene, casting long, dramatic shadows of deep cobalt blue, sculpting the contours of the terrain and imbuing the windswept snow with texture. The surface shimmers with millions of crystalline highlights, while the edges of the track catch bright light, contrasting sharply with the shadowy recesses. The camera's leisurely, majestic movement across this vast, silent landscape emphasizes a profound sense of majestic solitude and serene beauty, creating a breathtaking atmosphere of quiet contemplation at the heart of nature's grandeur.", + "seed": 8831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_4/index_5832_seed_8831_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04962308155372739, + "vlm_logprob": -0.04962308155372739, + "vlm_yes_no_margin": 3.799999976158142, + "vlm_token_credit": -0.04962308155372739, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1562124341726303 + ], + "opposite_logprob_tokens": [ + -2.031212329864502 + ], + "logprob_reward": -0.1562124341726303, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1562124341726303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25748106837272644, + 0.11467576026916504, + 0.07105793803930283, + 0.03617849946022034, + 0.08585754781961441, + 0.1076914444565773, + 0.1592082977294922, + 0.06567350029945374, + 0.06098020821809769, + 0.04119577258825302 + ], + "credit_map": [ + 0.011308400891721249, + 0.035027287900447845, + 0.032609570771455765, + 0.05574260279536247, + 0.02495597116649151, + 0.021191664040088654, + 0.04309295117855072, + 0.03355260565876961, + 0.008430317044258118, + 0.010048930533230305, + 0.008742338046431541, + 0.019628826528787613, + 0.007801812142133713, + 0.010429435409605503, + 0.02696659415960312, + 0.022627493366599083, + 0.008306342177093029, + 0.008372912183403969, + 0.010094965808093548, + 0.013606815598905087, + 0.005814795847982168, + 0.007144131697714329, + 0.009598170407116413, + 0.008119791746139526, + 0.002392540220171213, + 0.002409080509096384, + 0.004883900284767151, + 0.007483600173145533, + 0.003076087450608611, + 0.0038004762027412653, + 0.005964786745607853, + 0.006168024148792028, + 0.0037474280688911676, + 0.005749064031988382, + 0.0069307852536439896, + 0.021936431527137756, + 0.009504330344498158, + 0.008464501239359379, + 0.01731002703309059, + 0.012214972637593746, + 0.006468515377491713, + 0.007521019782871008, + 0.013933531939983368, + 0.01744571328163147, + 0.006824787240475416, + 0.007534356322139502, + 0.024029046297073364, + 0.023934464901685715, + 0.006617882754653692, + 0.009949584491550922, + 0.017072726041078568, + 0.030774828046560287, + 0.022227579727768898, + 0.03619947284460068, + 0.022314781323075294, + 0.01405143178999424, + 0.006260185036808252, + 0.004517010413110256, + 0.007545038592070341, + 0.01327530201524496, + 0.005994918756186962, + 0.008659105747938156, + 0.010760970413684845, + 0.008660961873829365, + 0.004508479498326778, + 0.0038522782269865274, + 0.007084389217197895, + 0.009360913187265396, + 0.004336113575845957, + 0.006260266527533531, + 0.011749914847314358, + 0.0138278529047966, + 0.003000899450853467, + 0.0017677551368251443, + 0.0047377534210681915, + 0.00423339894041419, + 0.0024991510435938835, + 0.0030997840221971273, + 0.011483595706522465, + 0.010373430326581001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.1562124341726303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865795746445656 + ], + "opposite_logprob_tokens": [ + -3.0486578941345215 + ], + "logprob_reward": -0.04865795746445656, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865795746445656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19156034290790558, + 0.1271502524614334, + 0.07468139380216599, + 0.09094646573066711, + 0.10863503068685532, + 0.06895454972982407, + 0.13743990659713745, + 0.06844726949930191, + 0.08424676209688187, + 0.047938015311956406 + ], + "credit_map": [ + 0.013090824708342552, + 0.027780616655945778, + 0.0156683549284935, + 0.04386003315448761, + 0.019445864483714104, + 0.016910551115870476, + 0.03154108300805092, + 0.023263027891516685, + 0.008667893707752228, + 0.012838458642363548, + 0.01084130723029375, + 0.030608223751187325, + 0.009214624762535095, + 0.010603063739836216, + 0.026486538350582123, + 0.01789015159010887, + 0.007555435877293348, + 0.007386635988950729, + 0.007358119357377291, + 0.013865796849131584, + 0.005036006215959787, + 0.008896879851818085, + 0.011831514537334442, + 0.012751004658639431, + 0.00798801239579916, + 0.008567736484110355, + 0.010798538103699684, + 0.022915976122021675, + 0.007760276552289724, + 0.009141298942267895, + 0.012622283771634102, + 0.011152347549796104, + 0.004850502125918865, + 0.006228175014257431, + 0.011884777806699276, + 0.03557822108268738, + 0.011150180362164974, + 0.010005566291511059, + 0.018999600782990456, + 0.0099380137398839, + 0.003161940723657608, + 0.004209449514746666, + 0.0063494122587144375, + 0.010286937467753887, + 0.004250993020832539, + 0.006182587705552578, + 0.016217948868870735, + 0.018295275047421455, + 0.004728520289063454, + 0.007840363308787346, + 0.01036886963993311, + 0.027092041447758675, + 0.018593603745102882, + 0.030688218772411346, + 0.023911643773317337, + 0.0142166493460536, + 0.0029314470011740923, + 0.003875561524182558, + 0.00555999344214797, + 0.015076628886163235, + 0.008513283915817738, + 0.01358647271990776, + 0.0105333486571908, + 0.008370534516870975, + 0.00828616414219141, + 0.005215155892074108, + 0.008200405165553093, + 0.014462263323366642, + 0.004717132076621056, + 0.006524045951664448, + 0.02097255364060402, + 0.01586904749274254, + 0.003936136141419411, + 0.0026851037982851267, + 0.006795855704694986, + 0.006572719197720289, + 0.002969581400975585, + 0.004975078161805868, + 0.01119420025497675, + 0.008809340186417103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.04865795746445656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615511115640402 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615511115640402, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615511115640402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26717066764831543, + 0.10260288417339325, + 0.14017616212368011, + 0.04267466813325882, + 0.0937323123216629, + 0.15573136508464813, + 0.05243946239352226, + 0.0524476021528244, + 0.05625505745410919, + 0.0367698073387146 + ], + "credit_map": [ + 0.018063094466924667, + 0.0396219827234745, + 0.031406424939632416, + 0.05408124253153801, + 0.029229123145341873, + 0.02443554997444153, + 0.0412609800696373, + 0.02907227911055088, + 0.010084097273647785, + 0.016876796260476112, + 0.010161707177758217, + 0.022805599495768547, + 0.007647149730473757, + 0.009432647377252579, + 0.016934210434556007, + 0.008660675026476383, + 0.012248522602021694, + 0.01580340415239334, + 0.015451405197381973, + 0.019156470894813538, + 0.010692214593291283, + 0.016279110684990883, + 0.027073366567492485, + 0.02347167208790779, + 0.004716480616480112, + 0.004949143622070551, + 0.004407147876918316, + 0.007892755791544914, + 0.005050645209848881, + 0.005070055834949017, + 0.005890249274671078, + 0.004698190838098526, + 0.005904801655560732, + 0.005136255640536547, + 0.009542597457766533, + 0.01267233956605196, + 0.00699600949883461, + 0.008568343706429005, + 0.019393397495150566, + 0.025518570095300674, + 0.009299635887145996, + 0.013159872964024544, + 0.01701931096613407, + 0.02150270715355873, + 0.013779847882688046, + 0.013105516321957111, + 0.03491383045911789, + 0.03295063227415085, + 0.001978676300495863, + 0.002477508969604969, + 0.005810569040477276, + 0.009177950210869312, + 0.005965592805296183, + 0.009131623432040215, + 0.010087371803820133, + 0.007810171693563461, + 0.004533414263278246, + 0.003737489692866802, + 0.004278003238141537, + 0.009501947090029716, + 0.0032655978575348854, + 0.0031995249446481466, + 0.013639925047755241, + 0.010291698388755322, + 0.004384712781757116, + 0.0038271795492619276, + 0.010122108273208141, + 0.011027912609279156, + 0.004041717853397131, + 0.006446466315537691, + 0.010456057265400887, + 0.005948902573436499, + 0.0024886580649763346, + 0.0019285629969090223, + 0.00373194576241076, + 0.004296890459954739, + 0.003067109966650605, + 0.0045707691460847855, + 0.008837464265525341, + 0.00784840527921915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.003615511115640402 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948578007519245 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948578007519245, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948578007519245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3661658465862274, + 0.16947467625141144, + 0.16877096891403198, + 0.05005348473787308, + 0.0452575720846653, + 0.042370013892650604, + 0.03613179177045822, + 0.049073249101638794, + 0.02901521325111389, + 0.0436871275305748 + ], + "credit_map": [ + 0.02265077643096447, + 0.04378867894411087, + 0.03188119828701019, + 0.08891560137271881, + 0.030248990282416344, + 0.03185434639453888, + 0.07067791372537613, + 0.046148329973220825, + 0.009677011519670486, + 0.01277066208422184, + 0.021453257650136948, + 0.047693826258182526, + 0.01326350774616003, + 0.01845891959965229, + 0.025131305679678917, + 0.021026184782385826, + 0.009747142903506756, + 0.01816585473716259, + 0.018687820062041283, + 0.025064202025532722, + 0.011645173653960228, + 0.012016129679977894, + 0.04335431009531021, + 0.0300903357565403, + 0.0021220752969384193, + 0.00416493508964777, + 0.005436508450657129, + 0.015614926815032959, + 0.0045044878497719765, + 0.004751032683998346, + 0.008370519615709782, + 0.005089001730084419, + 0.001999227562919259, + 0.00291431974619627, + 0.004378085490316153, + 0.010726217180490494, + 0.004605313763022423, + 0.004314816556870937, + 0.008562849834561348, + 0.007756741717457771, + 0.0024224608205258846, + 0.00168759620282799, + 0.003197244368493557, + 0.006436677183955908, + 0.0022517747711390257, + 0.003402133472263813, + 0.01122297439724207, + 0.011749153025448322, + 0.00200268835760653, + 0.001535635208711028, + 0.0025155532639473677, + 0.004451843909919262, + 0.0029334858991205692, + 0.0035966283176094294, + 0.009918696247041225, + 0.009177261032164097, + 0.0035052828025072813, + 0.002882025670260191, + 0.003798458492383361, + 0.006648222450166941, + 0.002738861134275794, + 0.0028252345509827137, + 0.014371022582054138, + 0.012304139323532581, + 0.0027231588028371334, + 0.001467154361307621, + 0.0025634588673710823, + 0.003525354666635394, + 0.0014837449416518211, + 0.002032262971624732, + 0.008125492371618748, + 0.0070945885963737965, + 0.0023871471639722586, + 0.0019912768620997667, + 0.0051954989321529865, + 0.008323592133820057, + 0.0035951905883848667, + 0.008431815542280674, + 0.007532836403697729, + 0.006229768507182598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005948578007519245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368092700839043 + ], + "opposite_logprob_tokens": [ + -3.4086809158325195 + ], + "logprob_reward": -0.03368092700839043, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368092700839043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19578909873962402, + 0.0691240131855011, + 0.10561486333608627, + 0.07694290578365326, + 0.07651329040527344, + 0.10271698981523514, + 0.1463058888912201, + 0.13252557814121246, + 0.046391524374485016, + 0.04807579889893532 + ], + "credit_map": [ + 0.007818013429641724, + 0.01824268512427807, + 0.016032325103878975, + 0.026492474600672722, + 0.01718415506184101, + 0.014933525584638119, + 0.05122452974319458, + 0.04386141896247864, + 0.006263475865125656, + 0.009359079413115978, + 0.00669375853613019, + 0.015132282860577106, + 0.007291513495147228, + 0.006884866859763861, + 0.010035540908575058, + 0.007463498506695032, + 0.006400606594979763, + 0.009036429226398468, + 0.006921009626239538, + 0.021480582654476166, + 0.00633201701566577, + 0.008864345960319042, + 0.02579284831881523, + 0.020787032321095467, + 0.0029179961420595646, + 0.005904879886657, + 0.009243311360478401, + 0.013111432082951069, + 0.0032968572340905666, + 0.007330691907554865, + 0.01725732535123825, + 0.01788041554391384, + 0.008058913052082062, + 0.0067870947532355785, + 0.008751850575208664, + 0.015387964434921741, + 0.008927702903747559, + 0.009814955294132233, + 0.00989457592368126, + 0.008890241384506226, + 0.004965614061802626, + 0.006920209154486656, + 0.011441976763308048, + 0.02840924635529518, + 0.008798114955425262, + 0.01276588998734951, + 0.01684490591287613, + 0.012571045197546482, + 0.011594939045608044, + 0.007753376383334398, + 0.014307851903140545, + 0.01868077926337719, + 0.009586982429027557, + 0.012941748835146427, + 0.035724908113479614, + 0.03571530431509018, + 0.012069464661180973, + 0.012632977217435837, + 0.00857521127909422, + 0.018712785094976425, + 0.008292271755635738, + 0.010850761085748672, + 0.03386687859892845, + 0.027525240555405617, + 0.005078181624412537, + 0.002520691603422165, + 0.0038415754679590464, + 0.00566059211269021, + 0.002375819720327854, + 0.003970512188971043, + 0.011943455785512924, + 0.01100070122629404, + 0.007522858213633299, + 0.0028268895111978054, + 0.004872668534517288, + 0.009101773612201214, + 0.004451492801308632, + 0.007521405816078186, + 0.007102470379322767, + 0.0046762446872889996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.03368092700839043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a panoramic high-angle drone perspective, the camera slowly and steadily glides right, tracing the path of a lone ski run carved into a vast, pristine mountainside covered in deep, untouched powder snow. The track itself is a delicate serpentine line, a graceful S-curve along the steep slope, its two parallel grooves sharply defined against the otherwise pristine white canvas. The low-angle rays of the late afternoon sun illuminate the scene, casting long, dramatic shadows of deep cobalt blue, sculpting the contours of the terrain and imbuing the windswept snow with texture. The surface shimmers with millions of crystalline highlights, while the edges of the track catch bright light, contrasting sharply with the shadowy recesses. The camera's leisurely, majestic movement across this vast, silent landscape emphasizes a profound sense of majestic solitude and serene beauty, creating a breathtaking atmosphere of quiet contemplation at the heart of nature's grandeur.", + "seed": 8832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_4/index_5832_seed_8832_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03471830105409026, + "vlm_logprob": -0.03471830105409026, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.03471830105409026, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0686783641576767 + ], + "opposite_logprob_tokens": [ + -2.818678379058838 + ], + "logprob_reward": -0.0686783641576767, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0686783641576767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3557701110839844, + 0.08904480189085007, + 0.10869841277599335, + 0.09980210661888123, + 0.05054178461432457, + 0.02692204713821411, + 0.03722837567329407, + 0.040356870740652084, + 0.03082980029284954, + 0.16080568730831146 + ], + "credit_map": [ + 0.035063184797763824, + 0.05136771500110626, + 0.05120070278644562, + 0.06773755699396133, + 0.03859961777925491, + 0.0331980437040329, + 0.05035031586885452, + 0.028252985328435898, + 0.0069073219783604145, + 0.011200372129678726, + 0.013558988459408283, + 0.015386202372610569, + 0.010169066488742828, + 0.009492716751992702, + 0.015747569501399994, + 0.0065825628116726875, + 0.012599080801010132, + 0.013875305652618408, + 0.019373374059796333, + 0.018473366275429726, + 0.01178801991045475, + 0.01003230083733797, + 0.012853733263909817, + 0.009703227318823338, + 0.005566138308495283, + 0.007264365442097187, + 0.020896032452583313, + 0.015775764361023903, + 0.005527087487280369, + 0.008753173984587193, + 0.02897527627646923, + 0.007044263184070587, + 0.005447836592793465, + 0.005375928245484829, + 0.00794158224016428, + 0.00818779319524765, + 0.003907566890120506, + 0.005809837486594915, + 0.006399477832019329, + 0.007471764925867319, + 0.0037507289089262486, + 0.0038350692484527826, + 0.00380084291100502, + 0.00362400826998055, + 0.002827107673510909, + 0.0033025266602635384, + 0.0029381546191871166, + 0.0028436090797185898, + 0.005538526456803083, + 0.005882055498659611, + 0.00591654609888792, + 0.006364159751683474, + 0.0035093924961984158, + 0.003052160143852234, + 0.0036298090126365423, + 0.0033357262145727873, + 0.005338410381227732, + 0.006267072632908821, + 0.0057923151180148125, + 0.007236320059746504, + 0.003481068182736635, + 0.0030708767008036375, + 0.004413514398038387, + 0.004757290706038475, + 0.0038844337686896324, + 0.003734411671757698, + 0.004225166514515877, + 0.005245475564152002, + 0.004100192803889513, + 0.003245995379984379, + 0.003465950256213546, + 0.0029281743336468935, + 0.01027025654911995, + 0.006257461383938789, + 0.05766275152564049, + 0.028496265411376953, + 0.01102210208773613, + 0.010561546310782433, + 0.009462427347898483, + 0.027072886005043983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0686783641576767 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980804070830345 + ], + "opposite_logprob_tokens": [ + -3.5298080444335938 + ], + "logprob_reward": -0.02980804070830345, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980804070830345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11767789721488953, + 0.29052144289016724, + 0.06654785573482513, + 0.20848627388477325, + 0.056548889726400375, + 0.050432927906513214, + 0.024344276636838913, + 0.024049848318099976, + 0.0270418431609869, + 0.13434873521327972 + ], + "credit_map": [ + 0.01297365315258503, + 0.017811676487326622, + 0.013522885739803314, + 0.019255030900239944, + 0.012542988173663616, + 0.01042958628386259, + 0.01881704293191433, + 0.012325024232268333, + 0.03477354347705841, + 0.04358366131782532, + 0.033734627068042755, + 0.04372217133641243, + 0.032868802547454834, + 0.038752008229494095, + 0.0352698378264904, + 0.027816804125905037, + 0.0059715635143220425, + 0.007432544603943825, + 0.013101333752274513, + 0.01159896794706583, + 0.00659588910639286, + 0.007731200195848942, + 0.006220246199518442, + 0.007896109484136105, + 0.019491542130708694, + 0.019036635756492615, + 0.02831076644361019, + 0.03030533716082573, + 0.01580936089158058, + 0.023288527503609657, + 0.039587151259183884, + 0.0326569490134716, + 0.0092230886220932, + 0.007538141217082739, + 0.006992957554757595, + 0.009151243604719639, + 0.0040781679563224316, + 0.00772538036108017, + 0.0054665496572852135, + 0.006373360753059387, + 0.004473173525184393, + 0.0075532859191298485, + 0.00510788056999445, + 0.007775211241096258, + 0.00415743188932538, + 0.006132610607892275, + 0.008502340875566006, + 0.006730992812663317, + 0.002969704568386078, + 0.0038663584273308516, + 0.0029854788444936275, + 0.003635401837527752, + 0.0027812751941382885, + 0.0022959141060709953, + 0.0028690178878605366, + 0.002941127633675933, + 0.003913862630724907, + 0.0035107319708913565, + 0.0030041357968002558, + 0.003544397186487913, + 0.0019656468648463488, + 0.0026889510918408632, + 0.0024789483286440372, + 0.002943175146356225, + 0.0032973976340144873, + 0.003152992110699415, + 0.0033892400097101927, + 0.004667626228183508, + 0.003074323059991002, + 0.0030254407320171595, + 0.0030742832459509373, + 0.003360540373250842, + 0.009301851503551006, + 0.007081959396600723, + 0.041108209639787674, + 0.022116031497716904, + 0.008234021253883839, + 0.008168638683855534, + 0.010881406255066395, + 0.0274566188454628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -0.02980804070830345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542534619569778 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542534619569778, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542534619569778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16406714916229248, + 0.10978984087705612, + 0.06421628594398499, + 0.10340854525566101, + 0.04984268918633461, + 0.027782386168837547, + 0.022913742810487747, + 0.02993684448301792, + 0.0913076400756836, + 0.33673495054244995 + ], + "credit_map": [ + 0.016401167958974838, + 0.024360770359635353, + 0.019271086901426315, + 0.032870303839445114, + 0.017037520185112953, + 0.017531922087073326, + 0.02531491592526436, + 0.011279460974037647, + 0.008970214053988457, + 0.02154991216957569, + 0.01623348519206047, + 0.02309388481080532, + 0.01040081586688757, + 0.008246474899351597, + 0.014506203122437, + 0.006788841914385557, + 0.0063812159933149815, + 0.006955287419259548, + 0.01176715549081564, + 0.01312317419797182, + 0.00600867485627532, + 0.005754232872277498, + 0.009702411480247974, + 0.00452412711456418, + 0.0108034061267972, + 0.009094939567148685, + 0.02297031320631504, + 0.014320469461381435, + 0.008111226372420788, + 0.010121438652276993, + 0.011168068274855614, + 0.016818678006529808, + 0.00615045428276062, + 0.004266063217073679, + 0.01007935032248497, + 0.006600131746381521, + 0.004181423224508762, + 0.005662799812853336, + 0.007630853448063135, + 0.005271610338240862, + 0.0027584428898990154, + 0.004312403500080109, + 0.003420579945668578, + 0.004864340182393789, + 0.002371271373704076, + 0.0035695231053978205, + 0.00380942621268332, + 0.0026763943023979664, + 0.0035854352172464132, + 0.0033808117732405663, + 0.003478624392300844, + 0.0037237463984638453, + 0.0019011966651305556, + 0.0019295811653137207, + 0.0027387221343815327, + 0.00217562192119658, + 0.004898641258478165, + 0.004166190046817064, + 0.004306046292185783, + 0.004203499760478735, + 0.003264613216742873, + 0.0030629246030002832, + 0.0030602309852838516, + 0.0029746973887085915, + 0.009186207316815853, + 0.008606837131083012, + 0.00841401144862175, + 0.009280865080654621, + 0.012162867933511734, + 0.007861185818910599, + 0.020040569826960564, + 0.01575508341193199, + 0.01475259568542242, + 0.003917749971151352, + 0.1412116289138794, + 0.03157214820384979, + 0.00959529634565115, + 0.01606646738946438, + 0.07997167855501175, + 0.03964734822511673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.012542534619569778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630723994225264 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630723994225264, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630723994225264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1728203296661377, + 0.18260186910629272, + 0.1111372858285904, + 0.11482326686382294, + 0.05524837225675583, + 0.03444549813866615, + 0.015195103362202644, + 0.02410019375383854, + 0.024411775171756744, + 0.26521629095077515 + ], + "credit_map": [ + 0.016123203560709953, + 0.023678915575146675, + 0.021673055365681648, + 0.02814754843711853, + 0.019157199189066887, + 0.020396927371621132, + 0.0271492637693882, + 0.016494236886501312, + 0.011918127536773682, + 0.020725293084979057, + 0.028462504968047142, + 0.03472275286912918, + 0.01827552169561386, + 0.019088203087449074, + 0.03279109671711922, + 0.01661839708685875, + 0.007083679083734751, + 0.00986025296151638, + 0.02498157136142254, + 0.022961199283599854, + 0.007451833691447973, + 0.012166453525424004, + 0.016019832342863083, + 0.010612472891807556, + 0.008389881812036037, + 0.013166802003979683, + 0.019772211089730263, + 0.026603471487760544, + 0.009433265775442123, + 0.01152108982205391, + 0.017414361238479614, + 0.008522198535501957, + 0.00896988995373249, + 0.007028081454336643, + 0.009385473094880581, + 0.008045694790780544, + 0.005098395049571991, + 0.005236429627984762, + 0.0069039021618664265, + 0.0045805140398442745, + 0.004541643429547548, + 0.0057076807133853436, + 0.0032689643558114767, + 0.0059189568273723125, + 0.00319641618989408, + 0.004270127974450588, + 0.004407172556966543, + 0.003134538419544697, + 0.0022759102284908295, + 0.0023338468745350838, + 0.002392022404819727, + 0.0027712909504771233, + 0.0012890278594568372, + 0.0013355596456676722, + 0.0013918319018557668, + 0.0014056166401132941, + 0.004636848345398903, + 0.003513603238388896, + 0.003163424553349614, + 0.0038023076485842466, + 0.002427252009510994, + 0.002343404572457075, + 0.0021706141997128725, + 0.002042740350589156, + 0.0035318259615451097, + 0.002460487885400653, + 0.002984421793371439, + 0.004086390137672424, + 0.00214062980376184, + 0.0019772504456341267, + 0.0036285310052335262, + 0.003602238604798913, + 0.011851545423269272, + 0.002948089735582471, + 0.11110720038414001, + 0.023996626958251, + 0.007140044122934341, + 0.012604711577296257, + 0.06471230089664459, + 0.030855786055326462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.007630723994225264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493184179067612 + ], + "opposite_logprob_tokens": [ + -2.929931879043579 + ], + "logprob_reward": -0.05493184179067612, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493184179067612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09210899472236633, + 0.045439377427101135, + 0.04895094037055969, + 0.04366437718272209, + 0.13464544713497162, + 0.20625384151935577, + 0.023211989551782608, + 0.053514305502176285, + 0.04271205887198448, + 0.3094986379146576 + ], + "credit_map": [ + 0.00854186899960041, + 0.009133102372288704, + 0.01731763780117035, + 0.01811208389699459, + 0.00854200404137373, + 0.01067381352186203, + 0.010905197821557522, + 0.008883286267518997, + 0.00684894435107708, + 0.006599476560950279, + 0.005545071791857481, + 0.006145777180790901, + 0.004211605526506901, + 0.005340260453522205, + 0.005028600804507732, + 0.00571964168921113, + 0.005080197472125292, + 0.006338132545351982, + 0.006685594562441111, + 0.01009258721023798, + 0.004376103635877371, + 0.005194507539272308, + 0.007015341892838478, + 0.004168475046753883, + 0.0038646382745355368, + 0.004018744453787804, + 0.011832626536488533, + 0.00600402383133769, + 0.0032460750080645084, + 0.0032632309012115, + 0.008396145887672901, + 0.0030388925224542618, + 0.02021956443786621, + 0.012981859035789967, + 0.02400798909366131, + 0.01591229997575283, + 0.009696464985609055, + 0.013334975577890873, + 0.023832669481635094, + 0.01465961430221796, + 0.020731093361973763, + 0.03802898898720741, + 0.024890370666980743, + 0.02958609163761139, + 0.02119327336549759, + 0.021281588822603226, + 0.03089938499033451, + 0.019643040373921394, + 0.0035762027837336063, + 0.0031461603939533234, + 0.0032048062421381474, + 0.003621648997068405, + 0.0023270389065146446, + 0.0024643042124807835, + 0.0022869312670081854, + 0.0025848979130387306, + 0.010566028766334057, + 0.007095667999237776, + 0.0074789272621273994, + 0.007231606170535088, + 0.006295158062130213, + 0.005657348781824112, + 0.004971339367330074, + 0.004218227695673704, + 0.004269830416887999, + 0.003321276744827628, + 0.007740168832242489, + 0.007433227263391018, + 0.0039252061396837234, + 0.0036400086246430874, + 0.007464794907718897, + 0.004917544778436422, + 0.012209155596792698, + 0.006982461549341679, + 0.10636633634567261, + 0.018822401762008667, + 0.020055308938026428, + 0.018296929076313972, + 0.1141471192240715, + 0.012618954293429852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.05493184179067612 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a panoramic high-angle drone perspective, the camera slowly and steadily glides right, tracing the path of a lone ski run carved into a vast, pristine mountainside covered in deep, untouched powder snow. The track itself is a delicate serpentine line, a graceful S-curve along the steep slope, its two parallel grooves sharply defined against the otherwise pristine white canvas. The low-angle rays of the late afternoon sun illuminate the scene, casting long, dramatic shadows of deep cobalt blue, sculpting the contours of the terrain and imbuing the windswept snow with texture. The surface shimmers with millions of crystalline highlights, while the edges of the track catch bright light, contrasting sharply with the shadowy recesses. The camera's leisurely, majestic movement across this vast, silent landscape emphasizes a profound sense of majestic solitude and serene beauty, creating a breathtaking atmosphere of quiet contemplation at the heart of nature's grandeur.", + "seed": 8833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_4/index_5832_seed_8833_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.053651089500635865, + "vlm_logprob": -0.053651089500635865, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.053651089500635865, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1695268154144287 + ], + "opposite_logprob_tokens": [ + -1.9195268154144287 + ], + "logprob_reward": -0.1695268154144287, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1695268154144287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11722982674837112, + 0.06465200334787369, + 0.253726065158844, + 0.1328023076057434, + 0.20496277511119843, + 0.05593346804380417, + 0.011659841053187847, + 0.103183314204216, + 0.036529913544654846, + 0.019320441409945488 + ], + "credit_map": [ + 0.013071687892079353, + 0.017429910600185394, + 0.01610303297638893, + 0.013816459104418755, + 0.015130000188946724, + 0.019313344731926918, + 0.00961502455174923, + 0.01275036670267582, + 0.00462197745218873, + 0.006936626508831978, + 0.0072400374338030815, + 0.005240355618298054, + 0.015197757631540298, + 0.01000993512570858, + 0.0036251626443117857, + 0.011780149303376675, + 0.009643333964049816, + 0.013431179337203503, + 0.010637912899255753, + 0.0055406647734344006, + 0.09845077991485596, + 0.07056371867656708, + 0.008525422774255276, + 0.03693303465843201, + 0.006556397303938866, + 0.007257387042045593, + 0.008679520338773727, + 0.0112227201461792, + 0.009701269678771496, + 0.03099653869867325, + 0.04309245944023132, + 0.015296019613742828, + 0.009093704633414745, + 0.00894547626376152, + 0.01426648534834385, + 0.011021275073289871, + 0.018863819539546967, + 0.048412173986434937, + 0.07821487635374069, + 0.016144976019859314, + 0.0023275765124708414, + 0.0028206477873027325, + 0.0035571714397519827, + 0.003400944173336029, + 0.003932514693588018, + 0.016217248514294624, + 0.02068515308201313, + 0.002992213238030672, + 0.0005798927741125226, + 0.0009842539438977838, + 0.0009723041439428926, + 0.0011103186989203095, + 0.0012198725016787648, + 0.0016853598644956946, + 0.0034369693603366613, + 0.0016708697658032179, + 0.004888836294412613, + 0.00687478668987751, + 0.006908492185175419, + 0.008490321226418018, + 0.008314619772136211, + 0.01153204683214426, + 0.035585977137088776, + 0.020588241517543793, + 0.0010967296548187733, + 0.0021168417297303677, + 0.0012910935329273343, + 0.0031765426974743605, + 0.012665539048612118, + 0.0025952414143830538, + 0.01087608840316534, + 0.0027118376456201077, + 0.001597995520569384, + 0.0019467638339847326, + 0.002782505704089999, + 0.004201978910714388, + 0.003382174763828516, + 0.001987870316952467, + 0.002204243326559663, + 0.0012169088004156947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.1695268154144287 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865773022174835 + ], + "opposite_logprob_tokens": [ + -3.0486576557159424 + ], + "logprob_reward": -0.04865773022174835, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865773022174835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11828108131885529, + 0.05449721962213516, + 0.15578651428222656, + 0.12954126298427582, + 0.12741321325302124, + 0.13473112881183624, + 0.06078283116221428, + 0.0591697134077549, + 0.11904837191104889, + 0.040748707950115204 + ], + "credit_map": [ + 0.009884349070489407, + 0.015677304938435555, + 0.014875267632305622, + 0.012221146374940872, + 0.015966929495334625, + 0.02159367874264717, + 0.011873397044837475, + 0.016189008951187134, + 0.004360707476735115, + 0.006503368727862835, + 0.006771284155547619, + 0.005120465997606516, + 0.009837464429438114, + 0.008226754143834114, + 0.004904373083263636, + 0.00877279881387949, + 0.008389429189264774, + 0.015383396297693253, + 0.011328025721013546, + 0.005822020582854748, + 0.0493476502597332, + 0.050113972276449203, + 0.006259961519390345, + 0.009142056107521057, + 0.004035583231598139, + 0.007561196107417345, + 0.007017228286713362, + 0.0070982445031404495, + 0.016199734061956406, + 0.04101036861538887, + 0.03816551715135574, + 0.008453394286334515, + 0.004689195193350315, + 0.0055414000526070595, + 0.008555794134736061, + 0.006537068169564009, + 0.014977348037064075, + 0.034067295491695404, + 0.04544278234243393, + 0.007602334022521973, + 0.003899591974914074, + 0.005203706678003073, + 0.007085314951837063, + 0.006354409269988537, + 0.01325219590216875, + 0.03413752093911171, + 0.05420016497373581, + 0.010598213411867619, + 0.003535948460921645, + 0.004621974192559719, + 0.0038193012587726116, + 0.004979202523827553, + 0.012077849358320236, + 0.0043386975303292274, + 0.02099481038749218, + 0.006415046285837889, + 0.0038733496330678463, + 0.005466242786496878, + 0.00439906632527709, + 0.005250503774732351, + 0.008856531232595444, + 0.008850950747728348, + 0.013860286213457584, + 0.008612786419689655, + 0.0039647589437663555, + 0.006671318784356117, + 0.004830735735595226, + 0.021116003394126892, + 0.04468206316232681, + 0.009754659608006477, + 0.020436596125364304, + 0.007592233829200268, + 0.0037105297669768333, + 0.00430880393832922, + 0.005693892482668161, + 0.00449172081425786, + 0.005536647979170084, + 0.004708438646048307, + 0.007295825518667698, + 0.005002846475690603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.04865773022174835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254150368273258 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254150368273258, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254150368273258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026007775217294693, + 0.0572403222322464, + 0.47791141271591187, + 0.1617070436477661, + 0.016246484592556953, + 0.09689535945653915, + 0.015885919332504272, + 0.08552016317844391, + 0.011002261191606522, + 0.05158326402306557 + ], + "credit_map": [ + 0.001526443287730217, + 0.003143659792840481, + 0.0020715470891445875, + 0.002214352134615183, + 0.0064623355865478516, + 0.005796146113425493, + 0.002240205882117152, + 0.002553085098043084, + 0.004390631802380085, + 0.006156061310321093, + 0.00646523293107748, + 0.004861857742071152, + 0.012885055504739285, + 0.008932030759751797, + 0.004155711270868778, + 0.009393744170665741, + 0.012268797494471073, + 0.016163483262062073, + 0.011559806764125824, + 0.007193485274910927, + 0.19846181571483612, + 0.1450185626745224, + 0.0155401099473238, + 0.07170534133911133, + 0.004235310014337301, + 0.006481204181909561, + 0.006099327001720667, + 0.004548646043986082, + 0.030857833102345467, + 0.028787488117814064, + 0.04107462987303734, + 0.03962259367108345, + 0.0009548827656544745, + 0.00100430054590106, + 0.0016437348676845431, + 0.0016549488063901663, + 0.0021294995676726103, + 0.0022860087919980288, + 0.00327616767026484, + 0.0032969412859529257, + 0.003808284644037485, + 0.005225309170782566, + 0.005713291000574827, + 0.00393159594386816, + 0.00934478547424078, + 0.021758178249001503, + 0.03661635145545006, + 0.0104975625872612, + 0.0007355556590482593, + 0.0010113362222909927, + 0.0011793867452070117, + 0.0012192658614367247, + 0.0024086907505989075, + 0.0019314616220071912, + 0.005599170457571745, + 0.0018010531784966588, + 0.004520750604569912, + 0.006358102895319462, + 0.005596626549959183, + 0.00781415868550539, + 0.010292787104845047, + 0.010949655435979366, + 0.024485060945153236, + 0.01550302654504776, + 0.0005894920323044062, + 0.001146090915426612, + 0.0008939780527725816, + 0.0010864693904295564, + 0.0011107635218650103, + 0.0013156712520867586, + 0.0028659659437835217, + 0.001993830781430006, + 0.0032013109885156155, + 0.0038636778481304646, + 0.00695444643497467, + 0.012601271271705627, + 0.004619876854121685, + 0.004574794787913561, + 0.008422194048762321, + 0.007345689460635185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.005254150368273258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736662704497576 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736662704497576, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736662704497576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048836905509233475, + 0.047204650938510895, + 0.3600991368293762, + 0.08214513212442398, + 0.11248758435249329, + 0.18091236054897308, + 0.03678886964917183, + 0.08906470239162445, + 0.021807100623846054, + 0.020653540268540382 + ], + "credit_map": [ + 0.002763162599876523, + 0.005028372630476952, + 0.004365629982203245, + 0.004502305760979652, + 0.01049997378140688, + 0.010827211663126945, + 0.004821620881557465, + 0.006028629839420319, + 0.0028920092154294252, + 0.003997860476374626, + 0.003784555708989501, + 0.0023591630160808563, + 0.013675978407263756, + 0.00775771914049983, + 0.0020906992722302675, + 0.010646664537489414, + 0.009329763241112232, + 0.011794847436249256, + 0.008219395764172077, + 0.005399818066507578, + 0.14903074502944946, + 0.10466024279594421, + 0.01207659114152193, + 0.059587717056274414, + 0.0018984557827934623, + 0.0020443096291273832, + 0.004096417222172022, + 0.002896336605772376, + 0.005629356950521469, + 0.01586332730948925, + 0.03510299697518349, + 0.014613927341997623, + 0.0035517816431820393, + 0.004374172072857618, + 0.007197026629000902, + 0.005994755774736404, + 0.010766115039587021, + 0.034654513001441956, + 0.04078107327222824, + 0.005168152041733265, + 0.007051319349557161, + 0.009509527124464512, + 0.01223661471158266, + 0.012314001098275185, + 0.01518612913787365, + 0.042814575135707855, + 0.065720334649086, + 0.0160798542201519, + 0.002336608013138175, + 0.0027528349310159683, + 0.0029203433077782393, + 0.002940974896773696, + 0.00648492481559515, + 0.0030954284593462944, + 0.01233630534261465, + 0.00392145337536931, + 0.0030588882509618998, + 0.004681306891143322, + 0.005026098806411028, + 0.007075109519064426, + 0.012369902804493904, + 0.007738957647234201, + 0.030081816017627716, + 0.01903262361884117, + 0.0012839920818805695, + 0.0020018797367811203, + 0.0017415378242731094, + 0.004547595977783203, + 0.006453337147831917, + 0.002374590840190649, + 0.0019562970846891403, + 0.0014478700468316674, + 0.0009282230166718364, + 0.0010418652091175318, + 0.0019308184273540974, + 0.007176448591053486, + 0.0030702869407832623, + 0.0014425592962652445, + 0.002953781047835946, + 0.0021095590200275183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.006736662704497576 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038080088794231415 + ], + "opposite_logprob_tokens": [ + -3.2880799770355225 + ], + "logprob_reward": -0.038080088794231415, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038080088794231415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04204006865620613, + 0.04907112568616867, + 0.1142594963312149, + 0.27454090118408203, + 0.161018967628479, + 0.1683451384305954, + 0.037419985979795456, + 0.08586287498474121, + 0.05127355456352234, + 0.016167880967259407 + ], + "credit_map": [ + 0.00432530278339982, + 0.006489777006208897, + 0.005123629234731197, + 0.00457122502848506, + 0.004647731315344572, + 0.00782692339271307, + 0.0039041773416101933, + 0.005151302553713322, + 0.0030659984331578016, + 0.004526298493146896, + 0.004642735235393047, + 0.0029175013769418, + 0.013446608558297157, + 0.00875846203416586, + 0.0020002038218080997, + 0.009713317267596722, + 0.004761932883411646, + 0.006532024592161179, + 0.006078910548239946, + 0.004398927558213472, + 0.04042835161089897, + 0.04240843653678894, + 0.005075238179415464, + 0.004575683735311031, + 0.008138990961015224, + 0.015113710425794125, + 0.011808337643742561, + 0.010687224566936493, + 0.060035739094018936, + 0.0731443390250206, + 0.038306355476379395, + 0.0573062039911747, + 0.008264866657555103, + 0.00752531411126256, + 0.013354253023862839, + 0.011304774321615696, + 0.014685546047985554, + 0.0347646027803421, + 0.058452412486076355, + 0.012667196802794933, + 0.008955731987953186, + 0.013133716769516468, + 0.016036905348300934, + 0.016483450308442116, + 0.016870690509676933, + 0.02607916295528412, + 0.04180913046002388, + 0.02897634543478489, + 0.0018871737411245704, + 0.002196674467995763, + 0.0024804221466183662, + 0.002672967268154025, + 0.007654301356524229, + 0.0032291400711983442, + 0.0138564957305789, + 0.003442808287218213, + 0.003618951654061675, + 0.00590429175645113, + 0.00616457499563694, + 0.008347844704985619, + 0.006416818127036095, + 0.007106160279363394, + 0.03055773302912712, + 0.01774650625884533, + 0.001954306848347187, + 0.003410488832741976, + 0.003853289410471916, + 0.005220247432589531, + 0.01699313148856163, + 0.004929357673972845, + 0.01206495612859726, + 0.0028477751184254885, + 0.0010799438459798694, + 0.001113506848923862, + 0.001874788897112012, + 0.004830340854823589, + 0.0024933626409620047, + 0.0013697493122890592, + 0.001692700432613492, + 0.0017134875524789095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.038080088794231415 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a panoramic high-angle drone perspective, the camera slowly and steadily glides right, tracing the path of a lone ski run carved into a vast, pristine mountainside covered in deep, untouched powder snow. The track itself is a delicate serpentine line, a graceful S-curve along the steep slope, its two parallel grooves sharply defined against the otherwise pristine white canvas. The low-angle rays of the late afternoon sun illuminate the scene, casting long, dramatic shadows of deep cobalt blue, sculpting the contours of the terrain and imbuing the windswept snow with texture. The surface shimmers with millions of crystalline highlights, while the edges of the track catch bright light, contrasting sharply with the shadowy recesses. The camera's leisurely, majestic movement across this vast, silent landscape emphasizes a profound sense of majestic solitude and serene beauty, creating a breathtaking atmosphere of quiet contemplation at the heart of nature's grandeur.", + "seed": 8834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_4/index_5832_seed_8834_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.21863511949777603, + "vlm_logprob": -0.21863511949777603, + "vlm_yes_no_margin": 1.725, + "vlm_token_credit": -0.21863511949777603, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6370441317558289 + ], + "opposite_logprob_tokens": [ + -0.7620441317558289 + ], + "logprob_reward": -0.6370441317558289, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6370441317558289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11479522287845612, + 0.10044607520103455, + 0.2728596031665802, + 0.015075596049427986, + 0.05489416420459747, + 0.36224502325057983, + 0.008838769048452377, + 0.022676924243569374, + 0.015377510339021683, + 0.03279109299182892 + ], + "credit_map": [ + 0.01507012452930212, + 0.01391609013080597, + 0.02141469530761242, + 0.014851431362330914, + 0.009569145739078522, + 0.0068621705286204815, + 0.011538567021489143, + 0.02157299965620041, + 0.010345444083213806, + 0.006326634902507067, + 0.01753273420035839, + 0.014123372733592987, + 0.006738715805113316, + 0.004912974312901497, + 0.025142930448055267, + 0.015323269180953503, + 0.011008104309439659, + 0.006320093758404255, + 0.09262676537036896, + 0.016500746831297874, + 0.007210275623947382, + 0.0045414576306939125, + 0.11351054906845093, + 0.02114161103963852, + 0.001573743298649788, + 0.0015346367144957185, + 0.001903191558085382, + 0.002284283982589841, + 0.0011725379154086113, + 0.00133478210773319, + 0.0024766409769654274, + 0.002795778913423419, + 0.003378996392711997, + 0.0026416529435664415, + 0.019790280610322952, + 0.00432892981916666, + 0.0036214624997228384, + 0.0035736667923629284, + 0.009054371155798435, + 0.00850480142980814, + 0.0109243243932724, + 0.011093437671661377, + 0.034314896911382675, + 0.0374477356672287, + 0.010166535153985023, + 0.0065034604631364346, + 0.15954506397247314, + 0.09224958717823029, + 0.0008238148293457925, + 0.0011178533313795924, + 0.0017671298701316118, + 0.001415870850905776, + 0.0005572473746724427, + 0.0005598410498350859, + 0.0014094437938183546, + 0.0011875675991177559, + 0.001283501973375678, + 0.0019772218074649572, + 0.003092796541750431, + 0.0033033511135727167, + 0.0010121694067493081, + 0.0008346590329892933, + 0.0074883317574858665, + 0.003684891387820244, + 0.0013449359685182571, + 0.0013568521244451404, + 0.0030670552514493465, + 0.002379459561780095, + 0.0014787637628614902, + 0.0007372271502390504, + 0.003256021998822689, + 0.0017571952193975449, + 0.0024430216290056705, + 0.003329014405608177, + 0.004410482011735439, + 0.005338379181921482, + 0.004004586022347212, + 0.002509835409000516, + 0.006413077935576439, + 0.004342698957771063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.6370441317558289 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904709666967392 + ], + "opposite_logprob_tokens": [ + -2.4640471935272217 + ], + "logprob_reward": -0.08904709666967392, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904709666967392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08947116136550903, + 0.19054552912712097, + 0.19872531294822693, + 0.035582274198532104, + 0.2508503794670105, + 0.0986536294221878, + 0.014796949923038483, + 0.025873931124806404, + 0.013155405409634113, + 0.0823453813791275 + ], + "credit_map": [ + 0.009291832335293293, + 0.008667395450174809, + 0.016609400510787964, + 0.018163906410336494, + 0.005020514130592346, + 0.0052292500622570515, + 0.012582811526954174, + 0.013906051404774189, + 0.01037270575761795, + 0.006062259431928396, + 0.025985442101955414, + 0.02067689597606659, + 0.006753843743354082, + 0.004077814985066652, + 0.07566563040018082, + 0.040950972586870193, + 0.007332200184464455, + 0.004851739387959242, + 0.057871703058481216, + 0.013238132931292057, + 0.005756719037890434, + 0.0026445987168699503, + 0.0857611745595932, + 0.0212690606713295, + 0.00473454175516963, + 0.003537969896569848, + 0.005721635650843382, + 0.004804547410458326, + 0.002142003271728754, + 0.002354805823415518, + 0.005757209379225969, + 0.006529560312628746, + 0.012186184525489807, + 0.0097073158249259, + 0.09408539533615112, + 0.014834391884505749, + 0.0145890973508358, + 0.011623458005487919, + 0.07582548260688782, + 0.017999060451984406, + 0.006099669262766838, + 0.007186773233115673, + 0.012193803675472736, + 0.010741990990936756, + 0.007221198175102472, + 0.004634639248251915, + 0.030196936801075935, + 0.02037862502038479, + 0.001698007807135582, + 0.0015425157034769654, + 0.0020903495606034994, + 0.0020896759815514088, + 0.000908444169908762, + 0.0009714648476801813, + 0.0032049654982984066, + 0.0022915268782526255, + 0.0021701373625546694, + 0.0032172072678804398, + 0.004300619475543499, + 0.003137144958600402, + 0.004155188798904419, + 0.003345161210745573, + 0.003062022617086768, + 0.002486451994627714, + 0.0012362555135041475, + 0.0017391284927725792, + 0.002078882185742259, + 0.001986689167097211, + 0.0017608476337045431, + 0.000736930756829679, + 0.0022527186665683985, + 0.0013639541575685143, + 0.01126457192003727, + 0.0065102181397378445, + 0.00900285504758358, + 0.00965452753007412, + 0.007305619306862354, + 0.005378534086048603, + 0.015462009236216545, + 0.017767054960131645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.08904709666967392 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295896768569946 + ], + "opposite_logprob_tokens": [ + -2.2379589080810547 + ], + "logprob_reward": -0.11295896768569946, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295896768569946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14999394118785858, + 0.22982743382453918, + 0.11715041846036911, + 0.03629589080810547, + 0.16803237795829773, + 0.07569129765033722, + 0.03266116976737976, + 0.07675769180059433, + 0.03368377685546875, + 0.07990594208240509 + ], + "credit_map": [ + 0.0135488361120224, + 0.022654486820101738, + 0.02322794869542122, + 0.021817734465003014, + 0.016154885292053223, + 0.01140802726149559, + 0.015520000830292702, + 0.025662025436758995, + 0.01364500354975462, + 0.009183099493384361, + 0.02693185582756996, + 0.030373619869351387, + 0.00910342950373888, + 0.005238947458565235, + 0.08689244836568832, + 0.04845903441309929, + 0.007959418930113316, + 0.0041487086564302444, + 0.04526137933135033, + 0.012715266086161137, + 0.0035958916414529085, + 0.003949773497879505, + 0.020149188116192818, + 0.019370796158909798, + 0.00423451280221343, + 0.0031611942686140537, + 0.008428286761045456, + 0.004322043154388666, + 0.0034870475064963102, + 0.002504599979147315, + 0.004497956018894911, + 0.005660250782966614, + 0.007535883225500584, + 0.006352245341986418, + 0.055046726018190384, + 0.01350303366780281, + 0.006776720285415649, + 0.0040356675162911415, + 0.06395617872476578, + 0.010825937613844872, + 0.005263757426291704, + 0.003783237189054489, + 0.010210921056568623, + 0.008906858041882515, + 0.0033088165801018476, + 0.0027304510585963726, + 0.02793322689831257, + 0.013554031029343605, + 0.0037246530409902334, + 0.002863602712750435, + 0.00613960949704051, + 0.006207559257745743, + 0.001978207379579544, + 0.001900024595670402, + 0.006466832477599382, + 0.003380679991096258, + 0.0044733271934092045, + 0.004775950685143471, + 0.011256156489253044, + 0.010092943906784058, + 0.003741195658221841, + 0.003914942033588886, + 0.02778835967183113, + 0.010714814998209476, + 0.0025104142259806395, + 0.002725637285038829, + 0.004842341877520084, + 0.003701356938108802, + 0.00338965211994946, + 0.0026470767334103584, + 0.007440747693181038, + 0.006426546722650528, + 0.007338357623666525, + 0.007197482977062464, + 0.00871215108782053, + 0.010740206576883793, + 0.00684604374691844, + 0.005166861228644848, + 0.021705636754631996, + 0.012199205346405506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.11295896768569946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270897388458252 + ], + "opposite_logprob_tokens": [ + -2.127089738845825 + ], + "logprob_reward": -0.1270897388458252, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270897388458252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1126134991645813, + 0.09258343279361725, + 0.10334745794534683, + 0.03078664280474186, + 0.043683771044015884, + 0.5204260349273682, + 0.028195273131132126, + 0.03191063553094864, + 0.011689556762576103, + 0.02476373128592968 + ], + "credit_map": [ + 0.010964770801365376, + 0.008344844914972782, + 0.025957593694329262, + 0.013062742538750172, + 0.005828662775456905, + 0.004395697731524706, + 0.01347692497074604, + 0.030582256615161896, + 0.009822512976825237, + 0.005058109760284424, + 0.010079614818096161, + 0.012016941793262959, + 0.005782187916338444, + 0.0035772782284766436, + 0.02902224473655224, + 0.017224537208676338, + 0.005653929430991411, + 0.002864296780899167, + 0.022998720407485962, + 0.00879175215959549, + 0.0041950359009206295, + 0.001996692968532443, + 0.04289357736706734, + 0.013953451998531818, + 0.0028783450834453106, + 0.002473660511896014, + 0.003901310497894883, + 0.0033429863397032022, + 0.0021382977720350027, + 0.0016232033958658576, + 0.005121467635035515, + 0.009307371452450752, + 0.002746743615716696, + 0.00281807710416615, + 0.006327204406261444, + 0.004640156868845224, + 0.003520478028804064, + 0.0022510753478854895, + 0.014190052635967731, + 0.007189983502030373, + 0.015709288418293, + 0.016222942620515823, + 0.050063591450452805, + 0.05227569490671158, + 0.014109100215137005, + 0.008989664725959301, + 0.22973819077014923, + 0.13331753015518188, + 0.0019897427409887314, + 0.00219975714571774, + 0.005884456913918257, + 0.0043913833796978, + 0.0022760715801268816, + 0.0012838803231716156, + 0.006731385365128517, + 0.0034385949838906527, + 0.0017445264384150505, + 0.0023160844575613737, + 0.004893751349300146, + 0.004550356417894363, + 0.0013277227990329266, + 0.0011697945883497596, + 0.010660055093467236, + 0.005248345900326967, + 0.0009083335171453655, + 0.0009617849718779325, + 0.0015129275852814317, + 0.001281239790841937, + 0.000999227981083095, + 0.0007567472057417035, + 0.002489558421075344, + 0.0027797373477369547, + 0.0021265025716274977, + 0.0024197010789066553, + 0.003213221440091729, + 0.003712656442075968, + 0.002738912822678685, + 0.0018493427196517587, + 0.005495438352227211, + 0.0032079543452709913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.1270897388458252 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera perspective a panoramic high-angle drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily glide to the right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the terrain a vast mountainside covered in deep, untouched powder snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ski track a delicate serpentine line with two parallel grooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are long shadows of deep cobalt blue visible on the snow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703566253185272 + ], + "opposite_logprob_tokens": [ + -2.127035617828369 + ], + "logprob_reward": -0.12703566253185272, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703566253185272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06782762706279755, + 0.03430485725402832, + 0.09722648561000824, + 0.020851388573646545, + 0.07621371746063232, + 0.43686234951019287, + 0.025404252111911774, + 0.043554823845624924, + 0.1133187785744667, + 0.08443575352430344 + ], + "credit_map": [ + 0.008861498907208443, + 0.007694033905863762, + 0.006200945004820824, + 0.009045843966305256, + 0.004443011712282896, + 0.003124925773590803, + 0.015064508654177189, + 0.013392857275903225, + 0.003115392057225108, + 0.0022245936561375856, + 0.002549119060859084, + 0.004897078964859247, + 0.002625030465424061, + 0.0022260781843215227, + 0.01037687063217163, + 0.006290695164352655, + 0.004227002616971731, + 0.0023607886396348476, + 0.03316764906048775, + 0.0053574503399431705, + 0.003015630878508091, + 0.0016560754738748074, + 0.03992081433534622, + 0.0075210812501609325, + 0.002658176003023982, + 0.0022166986018419266, + 0.003224832471460104, + 0.0033953937236219645, + 0.0018831078195944428, + 0.001197011792100966, + 0.0026901275850832462, + 0.0035860403440892696, + 0.004025774076581001, + 0.006033769343048334, + 0.016426509246230125, + 0.006407612934708595, + 0.004763658158481121, + 0.0041543953120708466, + 0.026222581043839455, + 0.008179422467947006, + 0.013825339265167713, + 0.013252342119812965, + 0.04218291863799095, + 0.04645107313990593, + 0.011651336215436459, + 0.007221390958875418, + 0.19377081096172333, + 0.10850714147090912, + 0.0019395335111767054, + 0.002124735387042165, + 0.00397613225504756, + 0.004771411884576082, + 0.001613326370716095, + 0.0013403005432337523, + 0.006118766963481903, + 0.003520045895129442, + 0.002443639561533928, + 0.003968673292547464, + 0.007501612417399883, + 0.0061884112656116486, + 0.0025346234906464815, + 0.0022474254947155714, + 0.01227178517729044, + 0.006398654542863369, + 0.006248627323657274, + 0.0056746224872767925, + 0.016774294897913933, + 0.013143305666744709, + 0.007391620427370071, + 0.006145417224615812, + 0.038323063403367996, + 0.019617831334471703, + 0.00617030868306756, + 0.009631936438381672, + 0.009603836573660374, + 0.010126921348273754, + 0.00946918223053217, + 0.008048196323215961, + 0.014101622626185417, + 0.01728374883532524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.12703566253185272 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_5/index_tensor([3357, 3357])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_5/index_tensor([3357, 3357])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3a6505badfef456bf0a47023bcc7124bc313a514 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_5/index_tensor([3357, 3357])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, menacing low-angle shot captures a grotesque, cube-shaped, gelatinous Hunter, suspended in abyssal, inky darkness. The entire scene is imbued with the vibrant, jagged energy of a Todd McFarlane painting, as dramatic backlighting from an unseen source carves the creature's menacing silhouette from the void. The Hunter's core is a translucent, bile-colored cube of jelly, its surface shimmering with moisture, its form subtly contracting and expanding with a life of its own. From its mass, a tangle of muscular tentacles erupts, each a black, veined muscle, taut and vibrating with contained energy, poised to strike decisively. These appendages are rendered in sharp, angular highlights and deep, inky shadows, culminating in carapace-like, razor-sharp barbs. The harsh backlight penetrates the edges of the gelatinous body, refracting through its dark interior to illuminate the swirling, blurred shapes within, creating a high-contrast, terrifying atmosphere of spiraling, imminent violence.", + "seed": 32578, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_5/index_3357_seed_32578_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05521929142996669, + "vlm_logprob": -0.05521929142996669, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.05521929142996669, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005318658426403999 + ], + "opposite_logprob_tokens": [ + -5.880318641662598 + ], + "logprob_reward": -0.005318658426403999, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.005318658426403999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08648718148469925, + 0.06747037172317505, + 0.07357058674097061, + 0.09223056584596634, + 0.05456383153796196, + 0.13621774315834045, + 0.06935598701238632, + 0.11272555589675903, + 0.1982450634241104, + 0.10913313180208206 + ], + "credit_map": [ + 0.007433087565004826, + 0.011233736760914326, + 0.007663346361368895, + 0.014612282626330853, + 0.005443147383630276, + 0.015188840217888355, + 0.006765404716134071, + 0.018147340044379234, + 0.005514303222298622, + 0.007310692220926285, + 0.00718477088958025, + 0.01004508975893259, + 0.005765503738075495, + 0.011244884692132473, + 0.004736878909170628, + 0.015668248757719994, + 0.007628291845321655, + 0.006852654740214348, + 0.00830242782831192, + 0.012301137670874596, + 0.00887841172516346, + 0.011840756051242352, + 0.005625185091048479, + 0.012141721323132515, + 0.009735353291034698, + 0.009538300335407257, + 0.010318372398614883, + 0.01515424158424139, + 0.012779117561876774, + 0.0120836878195405, + 0.00882728025317192, + 0.013794207014143467, + 0.005466558039188385, + 0.005870696622878313, + 0.005118937697261572, + 0.008908567950129509, + 0.0039894250221550465, + 0.008704966865479946, + 0.004564812872558832, + 0.011939863674342632, + 0.013127257116138935, + 0.01743975095450878, + 0.015971072018146515, + 0.02121543698012829, + 0.008968012407422066, + 0.017847426235675812, + 0.012264193966984749, + 0.029384596273303032, + 0.007538919802755117, + 0.014354128390550613, + 0.009114407934248447, + 0.008936063386499882, + 0.00444795424118638, + 0.009917070157825947, + 0.006751507055014372, + 0.008295938372612, + 0.011563466861844063, + 0.017047010362148285, + 0.011845956556499004, + 0.015307086519896984, + 0.0096516078338027, + 0.011476771906018257, + 0.016046879813075066, + 0.019786780700087547, + 0.023598507046699524, + 0.022651953622698784, + 0.026225639507174492, + 0.028744570910930634, + 0.015782659873366356, + 0.027959050610661507, + 0.0231665950268507, + 0.030116096138954163, + 0.01095378678292036, + 0.013654099777340889, + 0.019772225990891457, + 0.011481632478535175, + 0.009279933758080006, + 0.013297337107360363, + 0.017388300970196724, + 0.013305815868079662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.005318658426403999 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028335205279290676 + ], + "opposite_logprob_tokens": [ + -5.877833366394043 + ], + "logprob_reward": -0.0028335205279290676, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028335205279290676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08610419183969498, + 0.1316930502653122, + 0.10569653660058975, + 0.0885419026017189, + 0.04852459579706192, + 0.10195570439100266, + 0.07024184614419937, + 0.07601018995046616, + 0.14641299843788147, + 0.144818976521492 + ], + "credit_map": [ + 0.008387552574276924, + 0.009689869359135628, + 0.00999714806675911, + 0.013063628226518631, + 0.007902710698544979, + 0.013801916502416134, + 0.0097426762804389, + 0.013518690131604671, + 0.0070763626135885715, + 0.013070168904960155, + 0.014590462669730186, + 0.022571731358766556, + 0.008049778640270233, + 0.025397012010216713, + 0.007321979384869337, + 0.03361555188894272, + 0.00808346550911665, + 0.010508699342608452, + 0.0169901754707098, + 0.015874609351158142, + 0.008838281035423279, + 0.017136963084340096, + 0.009022878482937813, + 0.019241467118263245, + 0.007664852775633335, + 0.009571564383804798, + 0.009302378632128239, + 0.01676994375884533, + 0.012638161890208721, + 0.012342606671154499, + 0.007374613545835018, + 0.012877779081463814, + 0.006415325216948986, + 0.006727732717990875, + 0.004915677942335606, + 0.010602575726807117, + 0.0037649546284228563, + 0.007054142188280821, + 0.003169136354699731, + 0.00587505241855979, + 0.010300805792212486, + 0.011632680892944336, + 0.010742341168224812, + 0.016201268881559372, + 0.007941334508359432, + 0.014024913311004639, + 0.008176964707672596, + 0.022935397922992706, + 0.006654042750597, + 0.015145930461585522, + 0.007915609516203403, + 0.010574798099696636, + 0.0036183635238558054, + 0.010295269079506397, + 0.0073409550823271275, + 0.008696884848177433, + 0.007634292356669903, + 0.01316059473901987, + 0.007802590727806091, + 0.008585899136960506, + 0.005328975152224302, + 0.008832039311528206, + 0.00802650861442089, + 0.016639288514852524, + 0.019331390038132668, + 0.02187519334256649, + 0.015777116641402245, + 0.020477715879678726, + 0.013239267282187939, + 0.022569486871361732, + 0.013783417642116547, + 0.01935940608382225, + 0.01809685491025448, + 0.014890152029693127, + 0.023430952802300453, + 0.015538165345788002, + 0.014114427380263805, + 0.020299870520830154, + 0.01870976947247982, + 0.019738798961043358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.0028335205279290676 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864654965698719 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.00864654965698719, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864654965698719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07645715028047562, + 0.0657900795340538, + 0.037940144538879395, + 0.06446436047554016, + 0.060799043625593185, + 0.0883246511220932, + 0.07455319911241531, + 0.06077658012509346, + 0.25045549869537354, + 0.22043931484222412 + ], + "credit_map": [ + 0.0041396175511181355, + 0.008748323656618595, + 0.006081071682274342, + 0.011592418886721134, + 0.005700960289686918, + 0.015611080452799797, + 0.005382945761084557, + 0.019200729206204414, + 0.003629948478192091, + 0.00474125612527132, + 0.008815229870378971, + 0.010127948597073555, + 0.004777261056005955, + 0.012821502983570099, + 0.004404156468808651, + 0.016472775489091873, + 0.0040887207724153996, + 0.0037349199410527945, + 0.006161758676171303, + 0.005327282007783651, + 0.002248534467071295, + 0.007299198303371668, + 0.0017174866516143084, + 0.007362243719398975, + 0.005900290794670582, + 0.0085373530164361, + 0.008271599188446999, + 0.012413404881954193, + 0.005451552104204893, + 0.009383958764374256, + 0.004076005425304174, + 0.01043019164353609, + 0.00720747048035264, + 0.0077323573641479015, + 0.005610260181128979, + 0.01255794521421194, + 0.005785786546766758, + 0.010126306675374508, + 0.0042494554072618484, + 0.007529458962380886, + 0.008232544176280499, + 0.016544025391340256, + 0.008348310366272926, + 0.0164500642567873, + 0.00595984747633338, + 0.011700242757797241, + 0.009074217639863491, + 0.012015399523079395, + 0.007024932652711868, + 0.018335329368710518, + 0.007478996645659208, + 0.012954294681549072, + 0.004863696172833443, + 0.010075961239635944, + 0.0063177659176290035, + 0.007502223830670118, + 0.009483742527663708, + 0.0061685084365308285, + 0.008021567016839981, + 0.005887950770556927, + 0.004584372043609619, + 0.007852302864193916, + 0.007381228264421225, + 0.01139690913259983, + 0.025875505059957504, + 0.033845726400613785, + 0.02820577844977379, + 0.04206570237874985, + 0.01676286570727825, + 0.036227017641067505, + 0.031707026064395905, + 0.0357658788561821, + 0.02120099775493145, + 0.025710593909025192, + 0.03829021751880646, + 0.01915295049548149, + 0.018819602206349373, + 0.03439214453101158, + 0.030909672379493713, + 0.03196312487125397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.00864654965698719 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370213508605957 + ], + "opposite_logprob_tokens": [ + -3.4087021350860596 + ], + "logprob_reward": -0.03370213508605957, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370213508605957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12195479869842529, + 0.28573986887931824, + 0.19140854477882385, + 0.08082208037376404, + 0.04949992522597313, + 0.0711401030421257, + 0.01750476285815239, + 0.06076682358980179, + 0.06865174323320389, + 0.05251140892505646 + ], + "credit_map": [ + 0.007282687816768885, + 0.01272166520357132, + 0.011560223065316677, + 0.01652156561613083, + 0.009476946666836739, + 0.025617508217692375, + 0.007903226651251316, + 0.030870957300066948, + 0.016729649156332016, + 0.026550790295004845, + 0.0410088412463665, + 0.03885315731167793, + 0.021170923486351967, + 0.0538891963660717, + 0.017889106646180153, + 0.06964818388223648, + 0.01549378503113985, + 0.016231516376137733, + 0.03151370584964752, + 0.01932533271610737, + 0.016212671995162964, + 0.03825613856315613, + 0.00921971257776022, + 0.04515567421913147, + 0.006992947775870562, + 0.008069348521530628, + 0.010472528636455536, + 0.013806348666548729, + 0.009333419613540173, + 0.011751293204724789, + 0.005914152599871159, + 0.014482027851045132, + 0.005366786383092403, + 0.004186722449958324, + 0.006554658990353346, + 0.007680901326239109, + 0.004316125065088272, + 0.00821966864168644, + 0.004725868348032236, + 0.008449187502264977, + 0.008287467993795872, + 0.007308741565793753, + 0.008953164331614971, + 0.011269952170550823, + 0.006030770018696785, + 0.009299344383180141, + 0.0055054775439202785, + 0.01448518130928278, + 0.001868627266958356, + 0.0034975623711943626, + 0.0018371256301179528, + 0.003007239196449518, + 0.0011154217645525932, + 0.0021027368493378162, + 0.0020662806928157806, + 0.0020097678061574697, + 0.005982846952974796, + 0.008283409290015697, + 0.007907843217253685, + 0.008356040343642235, + 0.005321516189724207, + 0.005652448162436485, + 0.0097429808229208, + 0.009519727900624275, + 0.007547108922153711, + 0.010746687650680542, + 0.006435514893382788, + 0.01014703419059515, + 0.006977556273341179, + 0.009866610169410706, + 0.00924339797347784, + 0.00768782664090395, + 0.00657673692330718, + 0.005376458168029785, + 0.008832319639623165, + 0.00445215031504631, + 0.003919755108654499, + 0.009245284833014011, + 0.006639070343226194, + 0.007469627540558577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.03370213508605957 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255955934524536 + ], + "opposite_logprob_tokens": [ + -1.6005955934524536 + ], + "logprob_reward": -0.2255955934524536, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255955934524536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12081174552440643, + 0.06278911232948303, + 0.07855989784002304, + 0.12227203696966171, + 0.09209711104631424, + 0.20249980688095093, + 0.044418636709451675, + 0.049572862684726715, + 0.1447981297969818, + 0.08218067139387131 + ], + "credit_map": [ + 0.008385040797293186, + 0.011932517401874065, + 0.01442586351186037, + 0.01874520815908909, + 0.008801059797406197, + 0.024045394733548164, + 0.008019901812076569, + 0.02645675651729107, + 0.003463371191173792, + 0.006645237561315298, + 0.005284687504172325, + 0.008463018573820591, + 0.004184250719845295, + 0.012678729370236397, + 0.004032443277537823, + 0.018037371337413788, + 0.005335716530680656, + 0.011881327256560326, + 0.011314117349684238, + 0.01427367702126503, + 0.005655308719724417, + 0.012625910341739655, + 0.005006816238164902, + 0.012467021122574806, + 0.015485847368836403, + 0.01593288779258728, + 0.013464170508086681, + 0.024553095921874046, + 0.010851682163774967, + 0.017331307753920555, + 0.008853662759065628, + 0.0157993845641613, + 0.007721837610006332, + 0.014644846320152283, + 0.010296232998371124, + 0.011652939021587372, + 0.010211551561951637, + 0.01331628393381834, + 0.006360865198075771, + 0.01789255440235138, + 0.016647938638925552, + 0.03557369112968445, + 0.015910906717181206, + 0.03507636487483978, + 0.01711406745016575, + 0.026725519448518753, + 0.023938803002238274, + 0.03151252865791321, + 0.004048124421387911, + 0.010483999736607075, + 0.0038746215868741274, + 0.007369003724306822, + 0.0032112160697579384, + 0.005749630741775036, + 0.005091934464871883, + 0.004590103402733803, + 0.004439009819179773, + 0.005572972819209099, + 0.004635143093764782, + 0.007858007214963436, + 0.003789092181250453, + 0.0065711191855371, + 0.01021538209170103, + 0.006492133252322674, + 0.015260554850101471, + 0.023268690332770348, + 0.014139601029455662, + 0.025021493434906006, + 0.012842416763305664, + 0.016805127263069153, + 0.0203049648553133, + 0.017155295237898827, + 0.008202933706343174, + 0.008545086719095707, + 0.014243894256651402, + 0.011858507059514523, + 0.006465582177042961, + 0.014489790424704552, + 0.00915377028286457, + 0.009221111424267292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.2255955934524536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, menacing low-angle shot captures a grotesque, cube-shaped, gelatinous Hunter, suspended in abyssal, inky darkness. The entire scene is imbued with the vibrant, jagged energy of a Todd McFarlane painting, as dramatic backlighting from an unseen source carves the creature's menacing silhouette from the void. The Hunter's core is a translucent, bile-colored cube of jelly, its surface shimmering with moisture, its form subtly contracting and expanding with a life of its own. From its mass, a tangle of muscular tentacles erupts, each a black, veined muscle, taut and vibrating with contained energy, poised to strike decisively. These appendages are rendered in sharp, angular highlights and deep, inky shadows, culminating in carapace-like, razor-sharp barbs. The harsh backlight penetrates the edges of the gelatinous body, refracting through its dark interior to illuminate the swirling, blurred shapes within, creating a high-contrast, terrifying atmosphere of spiraling, imminent violence.", + "seed": 32579, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_5/index_3357_seed_32579_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03636805089190602, + "vlm_logprob": -0.03636805089190602, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.03636805089190602, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006111031398177147 + ], + "opposite_logprob_tokens": [ + -5.631111145019531 + ], + "logprob_reward": -0.006111031398177147, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.006111031398177147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18752257525920868, + 0.06760753691196442, + 0.046361371874809265, + 0.08627130836248398, + 0.055004168301820755, + 0.12131725251674652, + 0.06475330144166946, + 0.03644174709916115, + 0.10897529870271683, + 0.22574543952941895 + ], + "credit_map": [ + 0.01391338836401701, + 0.018637103959918022, + 0.017074182629585266, + 0.03010679967701435, + 0.02410612441599369, + 0.029013486579060555, + 0.022153040394186974, + 0.032518453896045685, + 0.005739476066082716, + 0.004626268520951271, + 0.006486393511295319, + 0.008214263245463371, + 0.008645269088447094, + 0.013586921617388725, + 0.009254972450435162, + 0.011053972877562046, + 0.004825341049581766, + 0.0029867622070014477, + 0.00555051164701581, + 0.00414084829390049, + 0.005694972816854715, + 0.010381449945271015, + 0.005752851255238056, + 0.007028635591268539, + 0.008219747804105282, + 0.005382823757827282, + 0.010277862660586834, + 0.006582613568753004, + 0.02144804410636425, + 0.014682979322969913, + 0.009229978546500206, + 0.010447260923683643, + 0.00811454001814127, + 0.005277601070702076, + 0.0063054123893380165, + 0.0058676451444625854, + 0.0036204110365360975, + 0.015225491486489773, + 0.003917936235666275, + 0.006675126031041145, + 0.011704372242093086, + 0.014555557630956173, + 0.0149020841345191, + 0.025116415694355965, + 0.011101863346993923, + 0.014254002831876278, + 0.01375283021479845, + 0.015930112451314926, + 0.008578372187912464, + 0.005425466690212488, + 0.015290990471839905, + 0.013994399458169937, + 0.00361440796405077, + 0.0038127831649035215, + 0.006784671451896429, + 0.007252213079482317, + 0.004301192704588175, + 0.0037817698903381824, + 0.005355360452085733, + 0.005643146578222513, + 0.0040386090986430645, + 0.003549744375050068, + 0.006377015262842178, + 0.0033949106000363827, + 0.008590984158217907, + 0.008547830395400524, + 0.01852397434413433, + 0.018416058272123337, + 0.008697657845914364, + 0.011392948217689991, + 0.01835045777261257, + 0.016455387696623802, + 0.018359476700425148, + 0.029327213764190674, + 0.032866913825273514, + 0.026180842891335487, + 0.03221946582198143, + 0.035989757627248764, + 0.030175164341926575, + 0.020626608282327652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.006111031398177147 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00410511763766408 + ], + "opposite_logprob_tokens": [ + -5.504105091094971 + ], + "logprob_reward": -0.00410511763766408, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00410511763766408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12721897661685944, + 0.13112495839595795, + 0.0674125999212265, + 0.07544124871492386, + 0.06130850687623024, + 0.03342070430517197, + 0.064454086124897, + 0.030337641015648842, + 0.11004604399204254, + 0.29923519492149353 + ], + "credit_map": [ + 0.01016615517437458, + 0.01295647956430912, + 0.00998846534639597, + 0.016200941056013107, + 0.021867288276553154, + 0.02229212410748005, + 0.012694817036390305, + 0.021052710711956024, + 0.00877669733017683, + 0.011505959555506706, + 0.011316085234284401, + 0.02115287445485592, + 0.019756991416215897, + 0.021494384855031967, + 0.017490003257989883, + 0.019631976261734962, + 0.0060286433435976505, + 0.004026139620691538, + 0.006291596218943596, + 0.007790182717144489, + 0.010070168413221836, + 0.01399221457540989, + 0.009473627433180809, + 0.009740032255649567, + 0.007385554723441601, + 0.006693010218441486, + 0.008528569713234901, + 0.007762606721371412, + 0.011268950067460537, + 0.014469929039478302, + 0.00890351366251707, + 0.01042911782860756, + 0.008725621737539768, + 0.009469022043049335, + 0.004905277397483587, + 0.005882232449948788, + 0.004620158579200506, + 0.015610896982252598, + 0.0047364868223667145, + 0.007358812261372805, + 0.005528751760721207, + 0.0042228903621435165, + 0.0042419396340847015, + 0.004636005032807589, + 0.002856367500498891, + 0.0031546393875032663, + 0.0037823133170604706, + 0.004997794050723314, + 0.008746064268052578, + 0.005139639601111412, + 0.015512239187955856, + 0.01332264393568039, + 0.004094114527106285, + 0.0039619444869458675, + 0.006450173910707235, + 0.007227263879030943, + 0.005796924699097872, + 0.0036564902402460575, + 0.004954480566084385, + 0.004430538974702358, + 0.002544674091041088, + 0.002965008607134223, + 0.002585426904261112, + 0.0034040973987430334, + 0.008105557411909103, + 0.009081955999135971, + 0.01793522946536541, + 0.01870334893465042, + 0.009143797680735588, + 0.011544357985258102, + 0.018708130344748497, + 0.016823671758174896, + 0.03331346809864044, + 0.03968977555632591, + 0.037862539291381836, + 0.031511276960372925, + 0.03860523924231529, + 0.047652266919612885, + 0.039721619337797165, + 0.030879007652401924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.00410511763766408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616936504840851 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.003616936504840851, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616936504840851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24091503024101257, + 0.09212638437747955, + 0.08321624249219894, + 0.08890311419963837, + 0.12162308394908905, + 0.0631844773888588, + 0.08523346483707428, + 0.05351812765002251, + 0.09467365592718124, + 0.0766064003109932 + ], + "credit_map": [ + 0.014823402278125286, + 0.020852509886026382, + 0.016424208879470825, + 0.026994390413165092, + 0.04583736136555672, + 0.04240819811820984, + 0.02678440511226654, + 0.04679053649306297, + 0.00839246902614832, + 0.0060759433545172215, + 0.007907758466899395, + 0.00928979180753231, + 0.015023971907794476, + 0.018233126029372215, + 0.012700393795967102, + 0.01450292393565178, + 0.008289830759167671, + 0.006083528511226177, + 0.0074472445994615555, + 0.010572763159871101, + 0.011879407800734043, + 0.016103992238640785, + 0.010917715728282928, + 0.011921758763492107, + 0.010302017442882061, + 0.0065441979095339775, + 0.008203119039535522, + 0.007488158997148275, + 0.015934709459543228, + 0.02085867151618004, + 0.00864015519618988, + 0.010932084172964096, + 0.017068112269043922, + 0.01449383981525898, + 0.01592998206615448, + 0.014544783160090446, + 0.011421197094023228, + 0.02033260464668274, + 0.010581238195300102, + 0.01725132204592228, + 0.00825676042586565, + 0.008093509823083878, + 0.007006527855992317, + 0.010122612118721008, + 0.006474572233855724, + 0.00910233985632658, + 0.006454371381551027, + 0.00767379067838192, + 0.008563678711652756, + 0.008763455785810947, + 0.014495332725346088, + 0.01410686131566763, + 0.0063226851634681225, + 0.007214369252324104, + 0.016120417043566704, + 0.009646673686802387, + 0.005469497758895159, + 0.00466240243986249, + 0.00794309377670288, + 0.009142703376710415, + 0.006876485887914896, + 0.006097070407122374, + 0.007098278030753136, + 0.006228594575077295, + 0.008474684320390224, + 0.011581797152757645, + 0.01653745211660862, + 0.010976684279739857, + 0.009222333319485188, + 0.017400650307536125, + 0.008267403580248356, + 0.012212649919092655, + 0.006338971666991711, + 0.009754995815455914, + 0.009787305258214474, + 0.008819324895739555, + 0.012960667721927166, + 0.011333286762237549, + 0.008961772546172142, + 0.008650070056319237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.003616936504840851 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076287128031253815 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.0076287128031253815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076287128031253815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1763543039560318, + 0.052603259682655334, + 0.03555206581950188, + 0.11889971047639847, + 0.14895200729370117, + 0.0791911855340004, + 0.2366965413093567, + 0.02613169141113758, + 0.08025436848402023, + 0.04536484181880951 + ], + "credit_map": [ + 0.009899713099002838, + 0.01195776741951704, + 0.014306709170341492, + 0.01972207985818386, + 0.03303837776184082, + 0.038599893450737, + 0.019163085147738457, + 0.02966669388115406, + 0.0036561088636517525, + 0.004271270241588354, + 0.0035371806006878614, + 0.006550306919962168, + 0.00789436511695385, + 0.01217798050493002, + 0.006508803926408291, + 0.008007252588868141, + 0.004303026013076305, + 0.0027256605681031942, + 0.0032305417116731405, + 0.0038029979914426804, + 0.004552602767944336, + 0.007216534111648798, + 0.005064790602773428, + 0.004655918106436729, + 0.009575202129781246, + 0.01166459545493126, + 0.014023481868207455, + 0.013312569819390774, + 0.02273646555840969, + 0.01687983237206936, + 0.014515628106892109, + 0.01619194820523262, + 0.02043035812675953, + 0.015040682628750801, + 0.01163645088672638, + 0.014964240603148937, + 0.014765040948987007, + 0.03616448864340782, + 0.011659843847155571, + 0.024290913715958595, + 0.009180757217109203, + 0.007310591172426939, + 0.014180407859385014, + 0.013885180465877056, + 0.004498396534472704, + 0.008130389265716076, + 0.007835409604012966, + 0.014170060865581036, + 0.03588421642780304, + 0.025672344490885735, + 0.04094620421528816, + 0.04094106703996658, + 0.017543716356158257, + 0.015892384573817253, + 0.03831405192613602, + 0.02150259166955948, + 0.002609844086691737, + 0.002573193982243538, + 0.004746775608509779, + 0.0063158017583191395, + 0.0015461095608770847, + 0.001337789581157267, + 0.0035164253786206245, + 0.0034857564605772495, + 0.007914547808468342, + 0.006198880262672901, + 0.013559951446950436, + 0.010174605064094067, + 0.006239194422960281, + 0.010497068986296654, + 0.011961833573877811, + 0.013708292506635189, + 0.005402057431638241, + 0.005469512660056353, + 0.006639428902417421, + 0.005088144913315773, + 0.007155665662139654, + 0.006681991741061211, + 0.005082451738417149, + 0.0038455966860055923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0076287128031253815 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037845611572266 + ], + "opposite_logprob_tokens": [ + -1.9103784561157227 + ], + "logprob_reward": -0.16037845611572266, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037845611572266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13948285579681396, + 0.028664832934737206, + 0.09009980410337448, + 0.08317114412784576, + 0.1179044246673584, + 0.04844097048044205, + 0.2296452820301056, + 0.07497640699148178, + 0.1460350751876831, + 0.041579216718673706 + ], + "credit_map": [ + 0.009490154683589935, + 0.012684506364166737, + 0.01071747113019228, + 0.01909598708152771, + 0.024123866111040115, + 0.019649110734462738, + 0.021314483135938644, + 0.022407282143831253, + 0.0024269430432468653, + 0.00207322696223855, + 0.0023721831385046244, + 0.004626118112355471, + 0.0031520668417215347, + 0.0058882213197648525, + 0.0033000526018440723, + 0.004826021380722523, + 0.01019260287284851, + 0.006502287928014994, + 0.007066582795232534, + 0.014163375832140446, + 0.008633265271782875, + 0.01105134654790163, + 0.020102355629205704, + 0.012387987226247787, + 0.00803580041974783, + 0.007748649455606937, + 0.007996913976967335, + 0.011141840368509293, + 0.01534460298717022, + 0.011015689931809902, + 0.010964617133140564, + 0.010923029854893684, + 0.018331272527575493, + 0.0067478385753929615, + 0.007469349540770054, + 0.007011473178863525, + 0.006677771918475628, + 0.047963544726371765, + 0.006967350374907255, + 0.016735829412937164, + 0.0044621676206588745, + 0.004389396868646145, + 0.006966364104300737, + 0.007964430376887321, + 0.004731542430818081, + 0.005547231994569302, + 0.006195057183504105, + 0.008184781298041344, + 0.030628349632024765, + 0.020841578021645546, + 0.05685001611709595, + 0.046484753489494324, + 0.01114264503121376, + 0.014140043407678604, + 0.027196643874049187, + 0.02236124873161316, + 0.00839140359312296, + 0.008505163714289665, + 0.0110789705067873, + 0.01443223562091589, + 0.005682178307324648, + 0.006476716138422489, + 0.01295583788305521, + 0.007453897502273321, + 0.012391279451549053, + 0.021736441180109978, + 0.0278383269906044, + 0.016246195882558823, + 0.01735696941614151, + 0.026508163660764694, + 0.01093737781047821, + 0.01302033569663763, + 0.004489177837967873, + 0.0057481396943330765, + 0.005915396846830845, + 0.00526475440710783, + 0.004389794077724218, + 0.0065508633852005005, + 0.005264996085315943, + 0.003956091124564409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.16037845611572266 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, menacing low-angle shot captures a grotesque, cube-shaped, gelatinous Hunter, suspended in abyssal, inky darkness. The entire scene is imbued with the vibrant, jagged energy of a Todd McFarlane painting, as dramatic backlighting from an unseen source carves the creature's menacing silhouette from the void. The Hunter's core is a translucent, bile-colored cube of jelly, its surface shimmering with moisture, its form subtly contracting and expanding with a life of its own. From its mass, a tangle of muscular tentacles erupts, each a black, veined muscle, taut and vibrating with contained energy, poised to strike decisively. These appendages are rendered in sharp, angular highlights and deep, inky shadows, culminating in carapace-like, razor-sharp barbs. The harsh backlight penetrates the edges of the gelatinous body, refracting through its dark interior to illuminate the swirling, blurred shapes within, creating a high-contrast, terrifying atmosphere of spiraling, imminent violence.", + "seed": 32580, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_5/index_3357_seed_32580_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05806971527636051, + "vlm_logprob": -0.05806971527636051, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.05806971527636051, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013411779887974262 + ], + "opposite_logprob_tokens": [ + -4.638411998748779 + ], + "logprob_reward": -0.013411779887974262, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013411779887974262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11167478561401367, + 0.09669367969036102, + 0.10058767348527908, + 0.22812211513519287, + 0.14947865903377533, + 0.05817694962024689, + 0.050049394369125366, + 0.046827442944049835, + 0.08415620028972626, + 0.07423307746648788 + ], + "credit_map": [ + 0.015021254308521748, + 0.010746690444648266, + 0.01644529588520527, + 0.017555754631757736, + 0.013559632934629917, + 0.01605275645852089, + 0.007828808389604092, + 0.014464598149061203, + 0.01103967521339655, + 0.00824564415961504, + 0.008880789391696453, + 0.01641724444925785, + 0.011897081509232521, + 0.013588672503829002, + 0.016248255968093872, + 0.010376316495239735, + 0.0078034172765910625, + 0.006796559318900108, + 0.010172145441174507, + 0.013704338110983372, + 0.011639718897640705, + 0.014323046430945396, + 0.02494599111378193, + 0.01120245736092329, + 0.02068784460425377, + 0.02045329473912716, + 0.026755448430776596, + 0.024886604398489, + 0.023007946088910103, + 0.02782512456178665, + 0.04931138455867767, + 0.03519447147846222, + 0.0187262874096632, + 0.015388778410851955, + 0.013550257310271263, + 0.019867412745952606, + 0.017645392566919327, + 0.01602334715425968, + 0.0281359925866127, + 0.02014118805527687, + 0.0069261775352060795, + 0.0074256970547139645, + 0.0073823933489620686, + 0.01002516783773899, + 0.003740753745660186, + 0.007078623399138451, + 0.007155830040574074, + 0.008442305028438568, + 0.005715612787753344, + 0.006369435228407383, + 0.00773930549621582, + 0.008082840591669083, + 0.004989273380488157, + 0.005062326788902283, + 0.005644354037940502, + 0.006446248386055231, + 0.006193611305207014, + 0.004841431975364685, + 0.006873005069792271, + 0.005946587771177292, + 0.004520769231021404, + 0.006110559683293104, + 0.0061401487328112125, + 0.006201330106705427, + 0.006211854983121157, + 0.006114328745752573, + 0.010786355473101139, + 0.010540483519434929, + 0.006862398702651262, + 0.009733742102980614, + 0.02171224355697632, + 0.012194789946079254, + 0.006296895910054445, + 0.006882955320179462, + 0.009309854358434677, + 0.009867360815405846, + 0.007733198814094067, + 0.008605374954640865, + 0.012942549772560596, + 0.012594891712069511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.013411779887974262 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012555248104035854 + ], + "opposite_logprob_tokens": [ + -4.387555122375488 + ], + "logprob_reward": -0.012555248104035854, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012555248104035854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18929477035999298, + 0.17016589641571045, + 0.15295691788196564, + 0.08111082017421722, + 0.11117664724588394, + 0.047379978001117706, + 0.03910541906952858, + 0.07604211568832397, + 0.08307173103094101, + 0.04969572648406029 + ], + "credit_map": [ + 0.0191117562353611, + 0.01815062202513218, + 0.02352979965507984, + 0.03850780799984932, + 0.016738034784793854, + 0.026665421202778816, + 0.02952008694410324, + 0.017071237787604332, + 0.015098239295184612, + 0.012598449364304543, + 0.014512516558170319, + 0.023693738505244255, + 0.016158007085323334, + 0.02437116578221321, + 0.04849575087428093, + 0.015238034538924694, + 0.011608751490712166, + 0.00960446149110794, + 0.01566017046570778, + 0.021928193047642708, + 0.01671367883682251, + 0.022395234555006027, + 0.04237888380885124, + 0.012667533941566944, + 0.007830692455172539, + 0.006736251059919596, + 0.010061991401016712, + 0.009340581484138966, + 0.0081217335537076, + 0.010710269212722778, + 0.017056094482541084, + 0.011253198608756065, + 0.013499279506504536, + 0.00928177498281002, + 0.00991943571716547, + 0.015297098085284233, + 0.014868713915348053, + 0.018150322139263153, + 0.01527602318674326, + 0.014883999712765217, + 0.005751962773501873, + 0.005983070004731417, + 0.006011266727000475, + 0.007762893568724394, + 0.003319879760965705, + 0.0053339758887887, + 0.006147978827357292, + 0.007068946957588196, + 0.003803983563557267, + 0.005212223157286644, + 0.005274459719657898, + 0.006705101113766432, + 0.003946714103221893, + 0.00418809661641717, + 0.005738012492656708, + 0.004236825276166201, + 0.009916081093251705, + 0.006938168779015541, + 0.008669192902743816, + 0.00990455225110054, + 0.008131307549774647, + 0.009243123233318329, + 0.011820299550890923, + 0.011419382877647877, + 0.0072977920062839985, + 0.004934146534651518, + 0.011026951484382153, + 0.008352281525731087, + 0.007985143922269344, + 0.010952390730381012, + 0.016156136989593506, + 0.016366885975003242, + 0.003811432281509042, + 0.0050650788471102715, + 0.0061564031057059765, + 0.005821658298373222, + 0.004904871340841055, + 0.0063548218458890915, + 0.00818544439971447, + 0.00939601194113493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012555248104035854 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197957426309586 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018197957426309586, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197957426309586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20060913264751434, + 0.10783563554286957, + 0.08151058852672577, + 0.11869847029447556, + 0.03696027770638466, + 0.19016247987747192, + 0.0556461326777935, + 0.10501086711883545, + 0.06192483380436897, + 0.04164159297943115 + ], + "credit_map": [ + 0.022022206336259842, + 0.028390806168317795, + 0.025313211604952812, + 0.03305545449256897, + 0.021131305024027824, + 0.02936432883143425, + 0.02165694534778595, + 0.019674865528941154, + 0.012513689696788788, + 0.009540612809360027, + 0.011622749269008636, + 0.015519535169005394, + 0.01255006343126297, + 0.01588432304561138, + 0.01993568055331707, + 0.010268984362483025, + 0.008577893488109112, + 0.005632159765809774, + 0.009717931039631367, + 0.011234921403229237, + 0.008542063646018505, + 0.012231690809130669, + 0.018967872485518456, + 0.0066060591489076614, + 0.01188509352505207, + 0.012243164703249931, + 0.01647687889635563, + 0.012089391238987446, + 0.01130736619234085, + 0.015575209632515907, + 0.020089885219931602, + 0.01903148740530014, + 0.006393176969140768, + 0.0037364063318818808, + 0.0026920735836029053, + 0.00506124272942543, + 0.00385965034365654, + 0.004789234604686499, + 0.003837654599919915, + 0.006590837147086859, + 0.026460405439138412, + 0.02328280359506607, + 0.021872499957680702, + 0.033508025109767914, + 0.01477479562163353, + 0.0218366589397192, + 0.0212919432669878, + 0.02713535726070404, + 0.006053858902305365, + 0.006750042084604502, + 0.00663199694827199, + 0.010311683639883995, + 0.00563087360933423, + 0.006675955839455128, + 0.008152026683092117, + 0.005439694970846176, + 0.013470740988850594, + 0.00955832377076149, + 0.012006852775812149, + 0.012577524408698082, + 0.010901996865868568, + 0.01213085651397705, + 0.016509661450982094, + 0.017854910343885422, + 0.006967827677726746, + 0.005418641958385706, + 0.008629456162452698, + 0.006983621045947075, + 0.006486673839390278, + 0.006973482668399811, + 0.010359146632254124, + 0.010105985216796398, + 0.002977003576233983, + 0.00479981629177928, + 0.004638726823031902, + 0.004755731206387281, + 0.005115865264087915, + 0.004838946275413036, + 0.00766972778365016, + 0.006845774129033089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.018197957426309586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581724122166634 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581724122166634, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581724122166634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15792129933834076, + 0.0712023377418518, + 0.0344666987657547, + 0.027070553973317146, + 0.09128531813621521, + 0.10276620090007782, + 0.05743435397744179, + 0.294385701417923, + 0.06139327585697174, + 0.10207431763410568 + ], + "credit_map": [ + 0.014113322831690311, + 0.01595998741686344, + 0.01981569267809391, + 0.031452201306819916, + 0.01432852353900671, + 0.021228428930044174, + 0.02670152299106121, + 0.014321626164019108, + 0.012678381986916065, + 0.0073660449124872684, + 0.009563799016177654, + 0.01001140009611845, + 0.006072323303669691, + 0.01233917847275734, + 0.006345836445689201, + 0.006825367454439402, + 0.0030800066888332367, + 0.0023875709157437086, + 0.004060602746903896, + 0.0050528449937701225, + 0.004258919972926378, + 0.005077540408819914, + 0.007874364033341408, + 0.0026748497039079666, + 0.002569627482444048, + 0.003296264447271824, + 0.0029167961329221725, + 0.003010497661307454, + 0.003450190182775259, + 0.0032542641274631023, + 0.0054526254534721375, + 0.003120287088677287, + 0.010731961578130722, + 0.008867492899298668, + 0.009183311834931374, + 0.011746781878173351, + 0.008458705618977547, + 0.012664551846683025, + 0.01333203911781311, + 0.016300469636917114, + 0.010723458603024483, + 0.012399411760270596, + 0.012919174507260323, + 0.016863344237208366, + 0.006477423943579197, + 0.011226614937186241, + 0.01667233183979988, + 0.015484446659684181, + 0.005459779407829046, + 0.004926350433379412, + 0.008374971337616444, + 0.007620856165885925, + 0.00530347740277648, + 0.006546730641275644, + 0.011308826506137848, + 0.007893363945186138, + 0.03691072016954422, + 0.022517075762152672, + 0.03866508975625038, + 0.029993686825037003, + 0.029601039364933968, + 0.03561360016465187, + 0.04763590544462204, + 0.05344857648015022, + 0.005240112077444792, + 0.004434372764080763, + 0.008749155327677727, + 0.007165888324379921, + 0.0044602639973163605, + 0.007167615462094545, + 0.014304609969258308, + 0.009871257469058037, + 0.009420191869139671, + 0.011134791187942028, + 0.01363098993897438, + 0.010909130796790123, + 0.008875248953700066, + 0.013596508651971817, + 0.012543591670691967, + 0.02196386270225048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.020581724122166634 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560186684131622 + ], + "opposite_logprob_tokens": [ + -1.6006019115447998 + ], + "logprob_reward": -0.22560186684131622, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560186684131622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19773496687412262, + 0.059058986604213715, + 0.05145607888698578, + 0.06730087846517563, + 0.09814276546239853, + 0.11484885215759277, + 0.191206157207489, + 0.08337423950433731, + 0.07451857626438141, + 0.062358491122722626 + ], + "credit_map": [ + 0.027784327045083046, + 0.01976877823472023, + 0.022348186001181602, + 0.026440907269716263, + 0.021061494946479797, + 0.02538054995238781, + 0.023233672603964806, + 0.031717054545879364, + 0.0076109678484499454, + 0.0047516291961073875, + 0.00882322434335947, + 0.009204810485243797, + 0.0059156110510230064, + 0.010720617137849331, + 0.00646610651165247, + 0.005566021893173456, + 0.006943256128579378, + 0.004899600986391306, + 0.008198501542210579, + 0.006193934474140406, + 0.006598302163183689, + 0.007808465976268053, + 0.005739316809922457, + 0.005074699874967337, + 0.008615683764219284, + 0.007439550012350082, + 0.010567286983132362, + 0.00915505550801754, + 0.006656886078417301, + 0.00970073975622654, + 0.008675720542669296, + 0.006489957682788372, + 0.019864186644554138, + 0.010081774555146694, + 0.011675682850182056, + 0.011492674238979816, + 0.008355147205293179, + 0.016662638634443283, + 0.009566030465066433, + 0.01044462900608778, + 0.015605307184159756, + 0.017605431377887726, + 0.01577703282237053, + 0.01953260973095894, + 0.008594946004450321, + 0.012004991993308067, + 0.012672873213887215, + 0.013055658899247646, + 0.025017814710736275, + 0.021988116204738617, + 0.02448456920683384, + 0.03170032426714897, + 0.017000064253807068, + 0.02026648446917534, + 0.029390262439846992, + 0.02135852724313736, + 0.01509107556194067, + 0.008672746829688549, + 0.009131976403295994, + 0.009701217524707317, + 0.010974406264722347, + 0.009544764645397663, + 0.010812308639287949, + 0.009445743635296822, + 0.007240853738039732, + 0.004853458143770695, + 0.00823503453284502, + 0.008595245890319347, + 0.009573040530085564, + 0.009516211226582527, + 0.012350364588201046, + 0.014154364354908466, + 0.005669215228408575, + 0.005788643844425678, + 0.005764390341937542, + 0.00842188112437725, + 0.009976312518119812, + 0.006340660620480776, + 0.011455245316028595, + 0.008942141197621822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.22560186684131622 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, menacing low-angle shot captures a grotesque, cube-shaped, gelatinous Hunter, suspended in abyssal, inky darkness. The entire scene is imbued with the vibrant, jagged energy of a Todd McFarlane painting, as dramatic backlighting from an unseen source carves the creature's menacing silhouette from the void. The Hunter's core is a translucent, bile-colored cube of jelly, its surface shimmering with moisture, its form subtly contracting and expanding with a life of its own. From its mass, a tangle of muscular tentacles erupts, each a black, veined muscle, taut and vibrating with contained energy, poised to strike decisively. These appendages are rendered in sharp, angular highlights and deep, inky shadows, culminating in carapace-like, razor-sharp barbs. The harsh backlight penetrates the edges of the gelatinous body, refracting through its dark interior to illuminate the swirling, blurred shapes within, creating a high-contrast, terrifying atmosphere of spiraling, imminent violence.", + "seed": 32581, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_5/index_3357_seed_32581_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03261545654386282, + "vlm_logprob": -0.03261545654386282, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.03261545654386282, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006049418356269598 + ], + "opposite_logprob_tokens": [ + -5.506049633026123 + ], + "logprob_reward": -0.006049418356269598, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006049418356269598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15092448890209198, + 0.04936929792165756, + 0.05871419608592987, + 0.16537240147590637, + 0.06980000436306, + 0.08112936466932297, + 0.015434276312589645, + 0.1676548719406128, + 0.18076753616333008, + 0.06083354353904724 + ], + "credit_map": [ + 0.015125330537557602, + 0.011957698501646519, + 0.027528801932930946, + 0.025289345532655716, + 0.010916163213551044, + 0.0216692965477705, + 0.01527029275894165, + 0.023167569190263748, + 0.006021967623382807, + 0.003975273575633764, + 0.006133731454610825, + 0.009412518702447414, + 0.003794279647991061, + 0.006614417303353548, + 0.008286734111607075, + 0.00513037433847785, + 0.005413725506514311, + 0.006556378211826086, + 0.006756103131920099, + 0.008775517344474792, + 0.00539738591760397, + 0.008721981197595596, + 0.009227875620126724, + 0.007865226827561855, + 0.011581695638597012, + 0.018116338178515434, + 0.024798694998025894, + 0.02020934224128723, + 0.019358115270733833, + 0.017528599128127098, + 0.021564224734902382, + 0.03221539780497551, + 0.00943069439381361, + 0.00468797143548727, + 0.023038458079099655, + 0.008676807396113873, + 0.011314470320940018, + 0.003856410039588809, + 0.0029044270049780607, + 0.005890759639441967, + 0.008077772334218025, + 0.004491779021918774, + 0.009419180452823639, + 0.00960009079426527, + 0.02230043150484562, + 0.008454863913357258, + 0.004864220507442951, + 0.013921020552515984, + 0.0020472314208745956, + 0.0018986398354172707, + 0.00197111489251256, + 0.002102021826431155, + 0.002268525306135416, + 0.0017254415433853865, + 0.0015737643698230386, + 0.0018475360702723265, + 0.010972618125379086, + 0.026860162615776062, + 0.005986726377159357, + 0.012319346889853477, + 0.054234493523836136, + 0.010578365065157413, + 0.014316265471279621, + 0.0323868952691555, + 0.013355507515370846, + 0.015613075345754623, + 0.00859586987644434, + 0.06077020987868309, + 0.012102596461772919, + 0.01407620869576931, + 0.010877082124352455, + 0.0453769825398922, + 0.011799700558185577, + 0.007988578639924526, + 0.005561931058764458, + 0.007844743318855762, + 0.006118090823292732, + 0.0054642255418002605, + 0.008946264162659645, + 0.007110010366886854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.006049418356269598 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028309053741395473 + ], + "opposite_logprob_tokens": [ + -5.877830982208252 + ], + "logprob_reward": -0.0028309053741395473, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028309053741395473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2489994913339615, + 0.09326334297657013, + 0.03421672806143761, + 0.09723906219005585, + 0.0424179770052433, + 0.05353163182735443, + 0.01978400908410549, + 0.3312327265739441, + 0.04532431438565254, + 0.03399071842432022 + ], + "credit_map": [ + 0.026562340557575226, + 0.019719626754522324, + 0.04418235272169113, + 0.046749528497457504, + 0.018137281760573387, + 0.034314874559640884, + 0.03392853960394859, + 0.02540493942797184, + 0.012126203626394272, + 0.006399079691618681, + 0.014638065360486507, + 0.015266739763319492, + 0.006381299812346697, + 0.011597656644880772, + 0.01474690716713667, + 0.01210738904774189, + 0.0032462170347571373, + 0.003858116688206792, + 0.004124798346310854, + 0.005475934129208326, + 0.003179574152454734, + 0.003987421747297049, + 0.005637099500745535, + 0.004707565065473318, + 0.007963063195347786, + 0.00823592022061348, + 0.015226348303258419, + 0.011039260774850845, + 0.011139645241200924, + 0.011758691631257534, + 0.010437213815748692, + 0.02143891714513302, + 0.005016586743295193, + 0.004569521639496088, + 0.006622627843171358, + 0.006862184498459101, + 0.008436142466962337, + 0.003534996882081032, + 0.0029454894829541445, + 0.004430430941283703, + 0.004620315507054329, + 0.0034183852840214968, + 0.006885444279760122, + 0.005847285967320204, + 0.014257978647947311, + 0.00722656212747097, + 0.00311783398501575, + 0.008157823234796524, + 0.0029194061644375324, + 0.0020151068456470966, + 0.0029104240238666534, + 0.0030971458181738853, + 0.0020269376691430807, + 0.002502647927030921, + 0.0017188548808917403, + 0.002593487035483122, + 0.02118133381009102, + 0.06406677514314651, + 0.00919085368514061, + 0.023015223443508148, + 0.11778997629880905, + 0.018601220101118088, + 0.022531768307089806, + 0.05485554412007332, + 0.005556127056479454, + 0.00812389887869358, + 0.002776459325104952, + 0.008058428764343262, + 0.004336772486567497, + 0.004949988331645727, + 0.004216095432639122, + 0.007306542247533798, + 0.006262726616114378, + 0.004694666247814894, + 0.0029537114314734936, + 0.0048862784169614315, + 0.0030794336926192045, + 0.0029718505684286356, + 0.0041442387737333775, + 0.0049978140741586685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0028309053741395473 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643016029149294 + ], + "opposite_logprob_tokens": [ + -5.379642963409424 + ], + "logprob_reward": -0.004643016029149294, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643016029149294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10455697029829025, + 0.05979905650019646, + 0.07536475360393524, + 0.26735106110572815, + 0.15202483534812927, + 0.05580412223935127, + 0.038828637450933456, + 0.062137335538864136, + 0.08458159863948822, + 0.09955157339572906 + ], + "credit_map": [ + 0.01222871895879507, + 0.008360973559319973, + 0.017696961760520935, + 0.02197306416928768, + 0.007452831603586674, + 0.015646640211343765, + 0.012212572619318962, + 0.008985208347439766, + 0.008120891638100147, + 0.00455115269869566, + 0.00954836793243885, + 0.010533950291574001, + 0.0041678184643387794, + 0.007910188287496567, + 0.008839280344545841, + 0.0061274063773453236, + 0.007149073760956526, + 0.00592126976698637, + 0.008614717982709408, + 0.008190657012164593, + 0.006600828841328621, + 0.012444581836462021, + 0.008256143890321255, + 0.01818748377263546, + 0.019711878150701523, + 0.029592182487249374, + 0.038282983005046844, + 0.031270582228899, + 0.030790245160460472, + 0.02966318652033806, + 0.038782741874456406, + 0.049257270991802216, + 0.018058082088828087, + 0.00891626626253128, + 0.013785451650619507, + 0.025298314169049263, + 0.0527782179415226, + 0.0119925020262599, + 0.011159072630107403, + 0.010036922991275787, + 0.005594331305474043, + 0.007657710928469896, + 0.010852343402802944, + 0.006073279306292534, + 0.007640998810529709, + 0.008452977985143661, + 0.003810873255133629, + 0.005721608176827431, + 0.005263167899101973, + 0.003476247424259782, + 0.005119465757161379, + 0.005578211508691311, + 0.004062896128743887, + 0.006530214101076126, + 0.0035384492948651314, + 0.00525998417288065, + 0.008475041016936302, + 0.011003963649272919, + 0.004310937598347664, + 0.007530579809099436, + 0.00872882828116417, + 0.005304690916091204, + 0.005799510050565004, + 0.01098378375172615, + 0.007620938587933779, + 0.008378996513783932, + 0.005340526811778545, + 0.0214072298258543, + 0.00876659620553255, + 0.009149588644504547, + 0.005630445666611195, + 0.01828727312386036, + 0.019182369112968445, + 0.01129677053540945, + 0.007689571939408779, + 0.01221093162894249, + 0.008395210839807987, + 0.014259294606745243, + 0.010888664051890373, + 0.015628760680556297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.004643016029149294 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735715549439192 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735715549439192, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735715549439192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3173925280570984, + 0.07862290740013123, + 0.0619688555598259, + 0.11441207677125931, + 0.08960175514221191, + 0.03985292837023735, + 0.013269009068608284, + 0.22124166786670685, + 0.028021762147545815, + 0.035616468638181686 + ], + "credit_map": [ + 0.0425918884575367, + 0.026082191616296768, + 0.053200043737888336, + 0.053655415773391724, + 0.020276257768273354, + 0.04734636843204498, + 0.03562004491686821, + 0.03862032666802406, + 0.008217085152864456, + 0.006853981874883175, + 0.01136590912938118, + 0.015701619908213615, + 0.0064101149328053, + 0.01126230787485838, + 0.010980824008584023, + 0.007831073366105556, + 0.006859967019408941, + 0.004649685230106115, + 0.007118322886526585, + 0.010267690755426884, + 0.005030490458011627, + 0.0076049757190048695, + 0.008138233795762062, + 0.0122994938865304, + 0.007918575778603554, + 0.012375767342746258, + 0.016095031052827835, + 0.014548595994710922, + 0.011166967451572418, + 0.015981221571564674, + 0.01730382815003395, + 0.019022095948457718, + 0.00797817949205637, + 0.005188578274101019, + 0.013138852082192898, + 0.014408974908292294, + 0.0315428264439106, + 0.006251192651689053, + 0.005775131285190582, + 0.005318016745150089, + 0.0038724993355572224, + 0.004521470982581377, + 0.004092348739504814, + 0.004939586855471134, + 0.0076791467145085335, + 0.004558435641229153, + 0.0029763304628431797, + 0.007213111966848373, + 0.0016986606642603874, + 0.001370898331515491, + 0.002047383924946189, + 0.0020733231212943792, + 0.001243160106241703, + 0.002220471389591694, + 0.0009099536691792309, + 0.0017051586182788014, + 0.014028203673660755, + 0.0361412949860096, + 0.007608333136886358, + 0.015067068859934807, + 0.0735207200050354, + 0.013111823238432407, + 0.017604777589440346, + 0.04415947198867798, + 0.0027751282323151827, + 0.003658342408016324, + 0.0016573375323787332, + 0.007285429164767265, + 0.0028275279328227043, + 0.002828956348821521, + 0.0023045723792165518, + 0.004684466868638992, + 0.0062679098919034, + 0.004958139266818762, + 0.00234793359413743, + 0.005044005811214447, + 0.0034658601507544518, + 0.0031975519377738237, + 0.005536970216780901, + 0.004798099398612976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.006735715549439192 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a grotesque, cube-shaped, gelatinous creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's core a translucent, bile-colored cube of jelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do black, veined muscular tentacles with razor-sharp barbs erupt from the creature's mass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature suspended in an abyssal, inky darkness environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a static, low-angle perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281822741031647 + ], + "opposite_logprob_tokens": [ + -2.0178182125091553 + ], + "logprob_reward": -0.14281822741031647, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281822741031647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19726787507534027, + 0.11563015729188919, + 0.05040423572063446, + 0.13475123047828674, + 0.27027249336242676, + 0.04319925233721733, + 0.030825955793261528, + 0.0366077795624733, + 0.08461560308933258, + 0.0364253930747509 + ], + "credit_map": [ + 0.027502216398715973, + 0.012516070157289505, + 0.0459144189953804, + 0.027598781511187553, + 0.00958594772964716, + 0.030441889539361, + 0.017300084233283997, + 0.02640845812857151, + 0.013328016735613346, + 0.01250649057328701, + 0.01793859712779522, + 0.01892281137406826, + 0.007063066586852074, + 0.017651556059718132, + 0.013265574350953102, + 0.01495405100286007, + 0.00468023493885994, + 0.004598979372531176, + 0.005289472173899412, + 0.008592143654823303, + 0.005259896162897348, + 0.005698763765394688, + 0.007477644365280867, + 0.008807102218270302, + 0.010062729939818382, + 0.010964193381369114, + 0.01898220181465149, + 0.01660861447453499, + 0.01547053549438715, + 0.016948681324720383, + 0.020154355093836784, + 0.025559915229678154, + 0.024730540812015533, + 0.014991572126746178, + 0.03896617144346237, + 0.0428987480700016, + 0.09614137560129166, + 0.018522581085562706, + 0.018228892236948013, + 0.015792611986398697, + 0.0038745065685361624, + 0.0034074035938829184, + 0.006959289312362671, + 0.0053984313271939754, + 0.006717954762279987, + 0.006201499607414007, + 0.0037081819027662277, + 0.006931982934474945, + 0.003386756172403693, + 0.0036906199529767036, + 0.004228582605719566, + 0.005738567095249891, + 0.0038305188063532114, + 0.004029189236462116, + 0.002529825083911419, + 0.003391895443201065, + 0.003959272522479296, + 0.00641342718154192, + 0.0029276260174810886, + 0.004924184642732143, + 0.004395593889057636, + 0.004869325086474419, + 0.004655038472265005, + 0.004463308956474066, + 0.005806186236441135, + 0.0072631980292499065, + 0.00383828766644001, + 0.026587752625346184, + 0.005026691127568483, + 0.007027395069599152, + 0.006473586894571781, + 0.02259250544011593, + 0.004124107304960489, + 0.003477684687823057, + 0.003568504471331835, + 0.004628291819244623, + 0.0032936674542725086, + 0.005769683513790369, + 0.005045241676270962, + 0.006518211215734482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.14281822741031647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_6/index_tensor([9052, 9052])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_6/index_tensor([9052, 9052])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b65e155e050ff2ce03ea80f79a2e8a1e0e101172 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_6/index_tensor([9052, 9052])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, top-down macro shot remains static, focused on a perfectly square block of cold, pale yellow butter in the center of a dark, matte cast iron skillet. Under the influence of intense, invisible heat from below, the butter's sharp edges instantly begin to soften and shimmer. A ring of transparent, shimmering liquid forms at its base, rapidly expanding as the solid structure inexorably collapses and collapses inward. The transformation accelerates, and the entire pie dissolves into a vibrant pool of seething, translucent golden liquid. Tiny, rippled bubbles erupt across the surface in a chaotic dance, creating a dynamic, fizzing texture that catches the bright, direct overhead light, giving the entire pool a hypnotic glow. The scene captures the swift and dramatic metamorphosis, creating a concentrated, intensely sensory atmosphere of pure thermal energy and culinary action.", + "seed": 10598, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_6/index_9052_seed_10598_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0432365067768842, + "vlm_logprob": -0.0432365067768842, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.0432365067768842, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024715732783079147 + ], + "opposite_logprob_tokens": [ + -5.524715900421143 + ], + "logprob_reward": -0.024715732783079147, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.024715732783079147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14101381599903107, + 0.0563010610640049, + 0.08032307773828506, + 0.23284569382667542, + 0.09433718025684357, + 0.06506993621587753, + 0.08503850549459457, + 0.06821741908788681, + 0.13715367019176483, + 0.03969963267445564 + ], + "credit_map": [ + 0.013569770380854607, + 0.018575605005025864, + 0.014421090483665466, + 0.03324573487043381, + 0.016013219952583313, + 0.014092711731791496, + 0.013932610861957073, + 0.017163068056106567, + 0.0052544171921908855, + 0.006147522944957018, + 0.00948643870651722, + 0.007278787437826395, + 0.005539220757782459, + 0.007407852448523045, + 0.00556067144498229, + 0.009626150131225586, + 0.009223321452736855, + 0.011403398588299751, + 0.00964901503175497, + 0.012612801976501942, + 0.015932857990264893, + 0.004977635573595762, + 0.007206920068711042, + 0.0093171251937747, + 0.02502281405031681, + 0.046272266656160355, + 0.028130576014518738, + 0.034954845905303955, + 0.03193371370434761, + 0.012648465111851692, + 0.03336064890027046, + 0.020522376522421837, + 0.006638448219746351, + 0.018580637872219086, + 0.008432441391050816, + 0.02039560116827488, + 0.011884976178407669, + 0.006508150137960911, + 0.01418230775743723, + 0.007714609149843454, + 0.00576053885743022, + 0.008987057022750378, + 0.0075815892778337, + 0.010015851818025112, + 0.010131223127245903, + 0.004382438492029905, + 0.01228104718029499, + 0.005930194165557623, + 0.004069475922733545, + 0.02220749855041504, + 0.007743561174720526, + 0.021228916943073273, + 0.007528092712163925, + 0.004310933873057365, + 0.010845165699720383, + 0.007104862947016954, + 0.005409203469753265, + 0.0160762220621109, + 0.006008520256727934, + 0.008824838325381279, + 0.011608805507421494, + 0.005915274377912283, + 0.009628376923501492, + 0.004746180959045887, + 0.0046553295105695724, + 0.050699494779109955, + 0.003063489915803075, + 0.010317650623619556, + 0.0076620085164904594, + 0.004412348847836256, + 0.05206220969557762, + 0.004281132481992245, + 0.003517626319080591, + 0.00859430618584156, + 0.0033806213177740574, + 0.006596898660063744, + 0.005554070230573416, + 0.004111316055059433, + 0.004880876746028662, + 0.003063918324187398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.024715732783079147 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005286760628223419 + ], + "opposite_logprob_tokens": [ + -5.255286693572998 + ], + "logprob_reward": -0.005286760628223419, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005286760628223419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09256386011838913, + 0.10632333904504776, + 0.08158872276544571, + 0.19003000855445862, + 0.10411142557859421, + 0.07172907143831253, + 0.1386895477771759, + 0.10856251418590546, + 0.08559214323759079, + 0.02080940268933773 + ], + "credit_map": [ + 0.009454205632209778, + 0.012108947150409222, + 0.009736764244735241, + 0.01749359630048275, + 0.011979635804891586, + 0.010888262651860714, + 0.009047229774296284, + 0.011855226941406727, + 0.009726684540510178, + 0.013647954910993576, + 0.013907630927860737, + 0.015320896171033382, + 0.01207831408828497, + 0.014082641340792179, + 0.010493770241737366, + 0.01706545054912567, + 0.00824679434299469, + 0.010018306784331799, + 0.009576565586030483, + 0.00980832427740097, + 0.013279246166348457, + 0.009295525029301643, + 0.009228783659636974, + 0.012135175988078117, + 0.014710593968629837, + 0.053871121257543564, + 0.019006971269845963, + 0.028939876705408096, + 0.033751633018255234, + 0.008501104079186916, + 0.01925802044570446, + 0.011990692466497421, + 0.0072096604853868484, + 0.02227448672056198, + 0.009113100357353687, + 0.02453429065644741, + 0.013341446407139301, + 0.006092176306992769, + 0.013683420605957508, + 0.007862844504415989, + 0.006856637541204691, + 0.013576212339103222, + 0.006831984966993332, + 0.009480093605816364, + 0.008716595359146595, + 0.006685663480311632, + 0.012320106849074364, + 0.0072617740370333195, + 0.005189619958400726, + 0.03793831169605255, + 0.009899306111037731, + 0.040298156440258026, + 0.012854388915002346, + 0.006960310507565737, + 0.018831057474017143, + 0.006718383636325598, + 0.01028453279286623, + 0.013851442374289036, + 0.011431090533733368, + 0.014363505877554417, + 0.02363545075058937, + 0.008526754565536976, + 0.0164023544639349, + 0.010067385621368885, + 0.0029115607030689716, + 0.027413493022322655, + 0.0029845593962818384, + 0.008408164605498314, + 0.005646745674312115, + 0.0030249168630689383, + 0.03186438977718353, + 0.003338312730193138, + 0.0013661953853443265, + 0.004360695835202932, + 0.0021919896826148033, + 0.003244988154619932, + 0.0032422011718153954, + 0.0015394783113151789, + 0.0032291067764163017, + 0.0016347479540854692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.005286760628223419 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331765554845333 + ], + "opposite_logprob_tokens": [ + -3.773317575454712 + ], + "logprob_reward": -0.02331765554845333, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331765554845333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4332294464111328, + 0.11973428726196289, + 0.027589276432991028, + 0.08129723370075226, + 0.0734170451760292, + 0.05842871218919754, + 0.11079678684473038, + 0.05449530482292175, + 0.01733778417110443, + 0.023674100637435913 + ], + "credit_map": [ + 0.047212015837430954, + 0.059435728937387466, + 0.04855944216251373, + 0.07997589558362961, + 0.05409138277173042, + 0.04821421951055527, + 0.03442579135298729, + 0.06131495162844658, + 0.009111407212913036, + 0.020076395943760872, + 0.01448142435401678, + 0.0171080082654953, + 0.014234964735805988, + 0.015570420771837234, + 0.012472016736865044, + 0.016679657623171806, + 0.002503030700609088, + 0.0034256477374583483, + 0.0038688580971211195, + 0.003929965663701296, + 0.0051014721393585205, + 0.0023891236633062363, + 0.003581862198188901, + 0.0027893176302313805, + 0.007424047216773033, + 0.019507497549057007, + 0.008577791042625904, + 0.012702626176178455, + 0.013869374990463257, + 0.004215987864881754, + 0.008936929516494274, + 0.006062980275601149, + 0.00575924851000309, + 0.012704293243587017, + 0.00830522459000349, + 0.013321056962013245, + 0.009044668637216091, + 0.005624071694910526, + 0.010751098394393921, + 0.007907381281256676, + 0.006229162681847811, + 0.006787748541682959, + 0.008740555495023727, + 0.007530802860856056, + 0.0049657681956887245, + 0.0069950297474861145, + 0.009717768989503384, + 0.007461873814463615, + 0.004511964041739702, + 0.03353652358055115, + 0.008497660979628563, + 0.032042328268289566, + 0.008287966251373291, + 0.005084428004920483, + 0.012151082046329975, + 0.006684829946607351, + 0.003398538799956441, + 0.007078823167830706, + 0.003283175639808178, + 0.006822768598794937, + 0.015694880858063698, + 0.003832463640719652, + 0.009807291440665722, + 0.0045773619785904884, + 0.0013075212482362986, + 0.00202868296764791, + 0.0021258466877043247, + 0.0022893587592989206, + 0.0028968167025595903, + 0.0020915460772812366, + 0.0030070061329752207, + 0.0015910060610622168, + 0.0015892820665612817, + 0.0032467031851410866, + 0.002032400807365775, + 0.0034715700894594193, + 0.004065395332872868, + 0.002108920132741332, + 0.004728836938738823, + 0.002430992666631937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02331765554845333 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024968183133751154 + ], + "opposite_logprob_tokens": [ + -6.002496719360352 + ], + "logprob_reward": -0.0024968183133751154, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024968183133751154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10387124121189117, + 0.058915406465530396, + 0.09830038994550705, + 0.11453767120838165, + 0.05426715314388275, + 0.09187508374452591, + 0.13433334231376648, + 0.06790179014205933, + 0.19885750114917755, + 0.07714039832353592 + ], + "credit_map": [ + 0.012161636725068092, + 0.010236354544758797, + 0.016125526279211044, + 0.01693287305533886, + 0.010753392241895199, + 0.011724445037543774, + 0.008672015741467476, + 0.017264997586607933, + 0.006316482555121183, + 0.005872379522770643, + 0.008852565661072731, + 0.008475764654576778, + 0.00709959352388978, + 0.005139932502061129, + 0.007809666451066732, + 0.009349021129310131, + 0.011533243581652641, + 0.014807196334004402, + 0.009462559595704079, + 0.01206341851502657, + 0.01741681806743145, + 0.0096919359639287, + 0.007867659442126751, + 0.01545756310224533, + 0.011844058521091938, + 0.023263996466994286, + 0.014136699959635735, + 0.017944008111953735, + 0.01674964651465416, + 0.0061702728271484375, + 0.015753014013171196, + 0.008675973862409592, + 0.004428611136972904, + 0.008083490654826164, + 0.006349815055727959, + 0.008900233544409275, + 0.007844807580113411, + 0.003608481725677848, + 0.009876505471765995, + 0.005175210069864988, + 0.006057390943169594, + 0.017754938453435898, + 0.007851817645132542, + 0.01712813228368759, + 0.015853967517614365, + 0.00673628319054842, + 0.013506880030035973, + 0.006985667627304792, + 0.0066960700787603855, + 0.030149336904287338, + 0.014938767068088055, + 0.031351275742053986, + 0.014108268544077873, + 0.007129163946956396, + 0.019950075075030327, + 0.010010399855673313, + 0.0048148371279239655, + 0.011011344380676746, + 0.005139916203916073, + 0.007877242751419544, + 0.01741793565452099, + 0.0065308380872011185, + 0.009063219651579857, + 0.006046449765563011, + 0.005776326172053814, + 0.06885642558336258, + 0.005610112566500902, + 0.01746896654367447, + 0.011972152628004551, + 0.005040890071541071, + 0.07830817997455597, + 0.0058244504034519196, + 0.005935990251600742, + 0.013921886682510376, + 0.007470258511602879, + 0.012117750942707062, + 0.012325660325586796, + 0.0057097384706139565, + 0.013640093617141247, + 0.006019022781401873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.0024968183133751154 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036556661128998 + ], + "opposite_logprob_tokens": [ + -1.9103655815124512 + ], + "logprob_reward": -0.16036556661128998, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036556661128998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17239101231098175, + 0.061627812683582306, + 0.13894416391849518, + 0.15199823677539825, + 0.11213769763708115, + 0.1023726537823677, + 0.10257035493850708, + 0.03754197433590889, + 0.0824480876326561, + 0.03796801716089249 + ], + "credit_map": [ + 0.018547797575592995, + 0.019302750006318092, + 0.02573843114078045, + 0.03916841745376587, + 0.014579755254089832, + 0.01818620041012764, + 0.011536319740116596, + 0.025331344455480576, + 0.006498449482023716, + 0.007111805025488138, + 0.01025412417948246, + 0.009654412977397442, + 0.005428292788565159, + 0.007075376342982054, + 0.006408088840544224, + 0.009197267703711987, + 0.016437558457255363, + 0.021162226796150208, + 0.013272009789943695, + 0.017816679552197456, + 0.021319925785064697, + 0.014201803132891655, + 0.012231557630002499, + 0.022502398118376732, + 0.014333087019622326, + 0.032089851796627045, + 0.015111804939806461, + 0.02427360601723194, + 0.02491684816777706, + 0.008028983138501644, + 0.018049445003271103, + 0.015194620005786419, + 0.007277999073266983, + 0.022967034950852394, + 0.011182727292180061, + 0.028167463839054108, + 0.013491020537912846, + 0.006907343864440918, + 0.013124086894094944, + 0.009020029567182064, + 0.007236997596919537, + 0.020043637603521347, + 0.009455226361751556, + 0.01990997977554798, + 0.01621006242930889, + 0.005726057104766369, + 0.016566714271903038, + 0.00722397631034255, + 0.0031020950991660357, + 0.039223622530698776, + 0.004413626156747341, + 0.03792190179228783, + 0.006107938475906849, + 0.002566483337432146, + 0.005661947652697563, + 0.003572738030925393, + 0.0035398241598159075, + 0.006722000893205404, + 0.003787021851167083, + 0.004282664507627487, + 0.006810438819229603, + 0.003573738969862461, + 0.005796291399747133, + 0.0030299904756247997, + 0.002358202589675784, + 0.02855219878256321, + 0.0023493359331041574, + 0.0074817826971411705, + 0.004472607281059027, + 0.0021538978908210993, + 0.032563336193561554, + 0.0025167313870042562, + 0.0026618472766131163, + 0.005149449687451124, + 0.004582551307976246, + 0.0050371428951621056, + 0.005391132086515427, + 0.003414012724533677, + 0.007375648245215416, + 0.004356234800070524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.16036556661128998 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, top-down macro shot remains static, focused on a perfectly square block of cold, pale yellow butter in the center of a dark, matte cast iron skillet. Under the influence of intense, invisible heat from below, the butter's sharp edges instantly begin to soften and shimmer. A ring of transparent, shimmering liquid forms at its base, rapidly expanding as the solid structure inexorably collapses and collapses inward. The transformation accelerates, and the entire pie dissolves into a vibrant pool of seething, translucent golden liquid. Tiny, rippled bubbles erupt across the surface in a chaotic dance, creating a dynamic, fizzing texture that catches the bright, direct overhead light, giving the entire pool a hypnotic glow. The scene captures the swift and dramatic metamorphosis, creating a concentrated, intensely sensory atmosphere of pure thermal energy and culinary action.", + "seed": 10599, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_6/index_9052_seed_10599_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12299125762656331, + "vlm_logprob": -0.12299125762656331, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.12299125762656331, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029282355681061745 + ], + "opposite_logprob_tokens": [ + -4.529282569885254 + ], + "logprob_reward": -0.029282355681061745, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.029282355681061745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2146368771791458, + 0.07310032844543457, + 0.12725506722927094, + 0.1258903443813324, + 0.04893682897090912, + 0.07126123458147049, + 0.06612948328256607, + 0.15118998289108276, + 0.10066326707601547, + 0.020936552435159683 + ], + "credit_map": [ + 0.03265576437115669, + 0.01192401722073555, + 0.0434703566133976, + 0.02252744510769844, + 0.01454675942659378, + 0.02908702939748764, + 0.023381082341074944, + 0.03704441711306572, + 0.01014676596969366, + 0.0031860778108239174, + 0.017870895564556122, + 0.005951157305389643, + 0.005969881545752287, + 0.012058595195412636, + 0.00690755108371377, + 0.01100939977914095, + 0.013329993933439255, + 0.009588840417563915, + 0.014936516992747784, + 0.011671556159853935, + 0.01006302796304226, + 0.03863757103681564, + 0.012393752112984657, + 0.016633814200758934, + 0.013864376582205296, + 0.006196242291480303, + 0.006901849526911974, + 0.018232436850667, + 0.007883355021476746, + 0.038789913058280945, + 0.016115911304950714, + 0.017906267195940018, + 0.004134251270443201, + 0.00442830566316843, + 0.005962238647043705, + 0.009586798027157784, + 0.004784023389220238, + 0.006514401640743017, + 0.005452714394778013, + 0.008074097335338593, + 0.005082482472062111, + 0.009409666061401367, + 0.00962364673614502, + 0.008634882979094982, + 0.010368538089096546, + 0.00791723933070898, + 0.013233519159257412, + 0.006991256959736347, + 0.004278403706848621, + 0.008222789503633976, + 0.008501168340444565, + 0.011141100898385048, + 0.011177152395248413, + 0.0066796233877539635, + 0.009478699415922165, + 0.006650550756603479, + 0.010520024225115776, + 0.019291112199425697, + 0.014850393868982792, + 0.02830694615840912, + 0.023259717971086502, + 0.0128794489428401, + 0.02537151612341404, + 0.016710827127099037, + 0.00622372142970562, + 0.013411817140877247, + 0.01206232514232397, + 0.026719966903328896, + 0.012704508379101753, + 0.005723637528717518, + 0.015438465401530266, + 0.00837883073836565, + 0.0016509492415934801, + 0.002431540982797742, + 0.0028349878266453743, + 0.0030157931614667177, + 0.0025944530498236418, + 0.0020396620966494083, + 0.0034605632536113262, + 0.0029086023569107056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.029282355681061745 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052775111980736256 + ], + "opposite_logprob_tokens": [ + -5.255277633666992 + ], + "logprob_reward": -0.0052775111980736256, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052775111980736256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2828051745891571, + 0.1301259845495224, + 0.05968382582068443, + 0.0997471809387207, + 0.20018890500068665, + 0.049463290721178055, + 0.06946372985839844, + 0.04721672832965851, + 0.03702973946928978, + 0.02427539974451065 + ], + "credit_map": [ + 0.0397743321955204, + 0.01783113367855549, + 0.055976130068302155, + 0.02550402097404003, + 0.022961458191275597, + 0.031186740845441818, + 0.03302032873034477, + 0.056551046669483185, + 0.016213320195674896, + 0.008514453656971455, + 0.022897718474268913, + 0.012386049143970013, + 0.015713652595877647, + 0.01551081147044897, + 0.01861404813826084, + 0.020275931805372238, + 0.0051916553638875484, + 0.005319360177963972, + 0.005646592006087303, + 0.007434885948896408, + 0.006426486652344465, + 0.007499329280108213, + 0.011916616931557655, + 0.010248898528516293, + 0.009870229288935661, + 0.007973633706569672, + 0.009496656246483326, + 0.01978406496345997, + 0.00890612043440342, + 0.01700223982334137, + 0.0095133688300848, + 0.01720087230205536, + 0.008222494274377823, + 0.004976840689778328, + 0.006460553966462612, + 0.007897433824837208, + 0.0066594951786100864, + 0.0786384642124176, + 0.011093128472566605, + 0.07624049484729767, + 0.004096957854926586, + 0.006985022686421871, + 0.005555788986384869, + 0.007460399996489286, + 0.007636347319930792, + 0.004329019226133823, + 0.007549170404672623, + 0.005850586108863354, + 0.004308952949941158, + 0.008387940935790539, + 0.006052929442375898, + 0.00892690196633339, + 0.012547352351248264, + 0.007002406287938356, + 0.012670746073126793, + 0.009566502645611763, + 0.0028414514381438494, + 0.0061184996739029884, + 0.004712152294814587, + 0.00661952281370759, + 0.009848283603787422, + 0.004141088109463453, + 0.009064018726348877, + 0.0038717121351510286, + 0.002430932130664587, + 0.004161399323493242, + 0.004430413246154785, + 0.004897060804069042, + 0.006064314860850573, + 0.0033907333854585886, + 0.00785091333091259, + 0.0038039733190089464, + 0.0013964507961645722, + 0.0034451878163963556, + 0.002147786319255829, + 0.00418531009927392, + 0.00454212399199605, + 0.0018544780323281884, + 0.004747976083308458, + 0.001956088235601783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0052775111980736256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330717444419861 + ], + "opposite_logprob_tokens": [ + -3.7733070850372314 + ], + "logprob_reward": -0.02330717444419861, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330717444419861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2856687903404236, + 0.120271697640419, + 0.11201641708612442, + 0.14244548976421356, + 0.06097579374909401, + 0.07865145057439804, + 0.05735126510262489, + 0.07891711592674255, + 0.03659043833613396, + 0.027111534029245377 + ], + "credit_map": [ + 0.04542626813054085, + 0.013382266275584698, + 0.0687744989991188, + 0.0292592141777277, + 0.01583198644220829, + 0.03713080286979675, + 0.034165967255830765, + 0.04169778525829315, + 0.019371725618839264, + 0.004255338571965694, + 0.04460037499666214, + 0.0067780292592942715, + 0.007122714072465897, + 0.012893276289105415, + 0.006437087897211313, + 0.018813150003552437, + 0.00907258503139019, + 0.00754943722859025, + 0.01253640279173851, + 0.009209757670760155, + 0.007969222962856293, + 0.03407460078597069, + 0.013211773708462715, + 0.01839263364672661, + 0.016024352982640266, + 0.00558509910479188, + 0.007172951940447092, + 0.01358755026012659, + 0.006326932925730944, + 0.05569787696003914, + 0.009289391338825226, + 0.028761325404047966, + 0.0024225867819041014, + 0.0032675035763531923, + 0.0037243934348225594, + 0.006337186321616173, + 0.005052025895565748, + 0.018885713070631027, + 0.004296202212572098, + 0.016990184783935547, + 0.005754868034273386, + 0.008538462221622467, + 0.010275189764797688, + 0.012533429078757763, + 0.010431394912302494, + 0.006501991301774979, + 0.016942434012889862, + 0.0076736765913665295, + 0.0035484882537275553, + 0.007321684155613184, + 0.008268516510725021, + 0.012435990385711193, + 0.006779169198125601, + 0.00518670491874218, + 0.009490742348134518, + 0.004319971892982721, + 0.005772794131189585, + 0.01052213367074728, + 0.007926049642264843, + 0.016577936708927155, + 0.010010430589318275, + 0.006906076800078154, + 0.013875680044293404, + 0.007326012011617422, + 0.0031894640997052193, + 0.003831131849437952, + 0.0039826929569244385, + 0.0036954237148165703, + 0.005946718156337738, + 0.0037659904919564724, + 0.006794355809688568, + 0.005384658928960562, + 0.0022248507011681795, + 0.0031107410322874784, + 0.004334570840001106, + 0.003751733573153615, + 0.003794650547206402, + 0.0031112388242036104, + 0.0038636389654129744, + 0.0029201090801507235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.02330717444419861 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371492773294449 + ], + "opposite_logprob_tokens": [ + -3.408715009689331 + ], + "logprob_reward": -0.03371492773294449, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371492773294449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11908003687858582, + 0.06588375568389893, + 0.09401046484708786, + 0.10554977506399155, + 0.13879458606243134, + 0.1381620615720749, + 0.061059463769197464, + 0.16826197504997253, + 0.09556243568658829, + 0.013635496608912945 + ], + "credit_map": [ + 0.01332928892225027, + 0.00855160690844059, + 0.0196028184145689, + 0.015742624178528786, + 0.007323199883103371, + 0.015508974902331829, + 0.012946865521371365, + 0.026074660941958427, + 0.009563873521983624, + 0.0028072695713490248, + 0.01818254590034485, + 0.00601009838283062, + 0.005000901408493519, + 0.008698791265487671, + 0.006410888396203518, + 0.009209389798343182, + 0.007959897629916668, + 0.007420148700475693, + 0.011233302764594555, + 0.00817143451422453, + 0.0062905168160796165, + 0.028700152412056923, + 0.010118174366652966, + 0.014116830192506313, + 0.010609921999275684, + 0.004164499696344137, + 0.006832461338490248, + 0.009333900175988674, + 0.00488568190485239, + 0.039012204855680466, + 0.010661078616976738, + 0.02005002088844776, + 0.007204119581729174, + 0.007450937759131193, + 0.01151736918836832, + 0.015721196308732033, + 0.009196450002491474, + 0.040218155831098557, + 0.009724514558911324, + 0.03776184096932411, + 0.010177799500524998, + 0.015642967075109482, + 0.016341906040906906, + 0.01731095463037491, + 0.025791820138692856, + 0.014908005483448505, + 0.023992782458662987, + 0.013995814137160778, + 0.0037342989817261696, + 0.007862080819904804, + 0.006239549722522497, + 0.008292792364954948, + 0.011168397963047028, + 0.0066176289692521095, + 0.009561154991388321, + 0.007583563216030598, + 0.012319464236497879, + 0.021019617095589638, + 0.021186547353863716, + 0.025065520778298378, + 0.027923783287405968, + 0.01671838015317917, + 0.028957443311810493, + 0.015071225352585316, + 0.0072080702520906925, + 0.011335653252899647, + 0.011960545554757118, + 0.014203004539012909, + 0.012746996246278286, + 0.009189830161631107, + 0.016179673373699188, + 0.012738657183945179, + 0.000898948113899678, + 0.001945761265233159, + 0.0014577157562598586, + 0.0020389275159686804, + 0.00245812744833529, + 0.0009804310975596309, + 0.0026153139770030975, + 0.0012402714928612113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.03371492773294449 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233743190765381 + ], + "opposite_logprob_tokens": [ + -0.8983743190765381 + ], + "logprob_reward": -0.5233743190765381, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233743190765381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24656447768211365, + 0.22191841900348663, + 0.07003609091043472, + 0.13600687682628632, + 0.07517781108617783, + 0.11501295864582062, + 0.03467836603522301, + 0.03666018694639206, + 0.025109760463237762, + 0.03883495554327965 + ], + "credit_map": [ + 0.03640149161219597, + 0.011984254233539104, + 0.05029856786131859, + 0.027715718373656273, + 0.01436923909932375, + 0.0360107384622097, + 0.030847571790218353, + 0.03893691301345825, + 0.02661783993244171, + 0.016163012012839317, + 0.0498960018157959, + 0.02758212201297283, + 0.01923529990017414, + 0.0242606271058321, + 0.02631291188299656, + 0.03185061737895012, + 0.0060968054458498955, + 0.008659587241709232, + 0.007775452919304371, + 0.00984289962798357, + 0.009733024053275585, + 0.007245897315442562, + 0.012241699732840061, + 0.008440740406513214, + 0.013308858498930931, + 0.007113905157893896, + 0.009211115539073944, + 0.02194216288626194, + 0.0076860422268509865, + 0.043422289192676544, + 0.011251348070800304, + 0.022071175277233124, + 0.004659648519009352, + 0.009433372877538204, + 0.008963542059063911, + 0.010952063836157322, + 0.009711873717606068, + 0.008774173445999622, + 0.007499811705201864, + 0.015183336101472378, + 0.008309667930006981, + 0.015536276623606682, + 0.016171036288142204, + 0.01747567392885685, + 0.016887914389371872, + 0.013586780987679958, + 0.016935216262936592, + 0.01011040061712265, + 0.0020321267656981945, + 0.004376612138003111, + 0.0029896809719502926, + 0.0040854825638234615, + 0.0066209821961820126, + 0.0038365127984434366, + 0.0061437226831912994, + 0.004593250807374716, + 0.002817461732774973, + 0.004697293043136597, + 0.005150154232978821, + 0.005453615915030241, + 0.005535734351724386, + 0.0036703809164464474, + 0.005508162546902895, + 0.0038273860700428486, + 0.001965428702533245, + 0.0027731384616345167, + 0.0030388825107365847, + 0.0029127364978194237, + 0.004090430680662394, + 0.002549579134210944, + 0.004855696111917496, + 0.0029238704591989517, + 0.0026585771702229977, + 0.004338355269283056, + 0.006455263122916222, + 0.006212018895894289, + 0.005378298927098513, + 0.0033952088560909033, + 0.007109117228537798, + 0.0032881225924938917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.5233743190765381 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, top-down macro shot remains static, focused on a perfectly square block of cold, pale yellow butter in the center of a dark, matte cast iron skillet. Under the influence of intense, invisible heat from below, the butter's sharp edges instantly begin to soften and shimmer. A ring of transparent, shimmering liquid forms at its base, rapidly expanding as the solid structure inexorably collapses and collapses inward. The transformation accelerates, and the entire pie dissolves into a vibrant pool of seething, translucent golden liquid. Tiny, rippled bubbles erupt across the surface in a chaotic dance, creating a dynamic, fizzing texture that catches the bright, direct overhead light, giving the entire pool a hypnotic glow. The scene captures the swift and dramatic metamorphosis, creating a concentrated, intensely sensory atmosphere of pure thermal energy and culinary action.", + "seed": 10600, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_6/index_9052_seed_10600_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10340401320718229, + "vlm_logprob": -0.10340401320718229, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.10340401320718229, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019487589597702026 + ], + "opposite_logprob_tokens": [ + -5.269487380981445 + ], + "logprob_reward": -0.019487589597702026, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.019487589597702026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07426001131534576, + 0.14696183800697327, + 0.04217081144452095, + 0.18957854807376862, + 0.04431312903761864, + 0.1391979157924652, + 0.09426043182611465, + 0.1083889827132225, + 0.06630194932222366, + 0.09456643462181091 + ], + "credit_map": [ + 0.008058620616793633, + 0.007401800714433193, + 0.010223424062132835, + 0.011832435615360737, + 0.007399861700832844, + 0.008194610476493835, + 0.009307028725743294, + 0.011842227540910244, + 0.02985250949859619, + 0.008124558255076408, + 0.01381288655102253, + 0.02048759162425995, + 0.008455285802483559, + 0.012355087324976921, + 0.013819772750139236, + 0.04005415365099907, + 0.005041706841439009, + 0.00501915393397212, + 0.00435460964217782, + 0.004979461431503296, + 0.004609814845025539, + 0.006209089420735836, + 0.0070633068680763245, + 0.0048936656676232815, + 0.010877280496060848, + 0.03287779912352562, + 0.009350255131721497, + 0.01824728585779667, + 0.05315075442194939, + 0.011569974943995476, + 0.024406785145401955, + 0.02909841202199459, + 0.004258220084011555, + 0.004424471408128738, + 0.0033893759828060865, + 0.007054481189697981, + 0.00493042916059494, + 0.003629629034548998, + 0.008682983927428722, + 0.007943539880216122, + 0.008099089376628399, + 0.030064698308706284, + 0.005098829977214336, + 0.011613814160227776, + 0.03137339651584625, + 0.007011965848505497, + 0.028263816609978676, + 0.017672304064035416, + 0.0031340960413217545, + 0.030321212485432625, + 0.004336741287261248, + 0.005560364108532667, + 0.02938457950949669, + 0.004854465834796429, + 0.011203994043171406, + 0.005464981775730848, + 0.007970881648361683, + 0.017441529780626297, + 0.013383187353610992, + 0.00967364851385355, + 0.015087416395545006, + 0.0070725297555327415, + 0.025920657441020012, + 0.011839139275252819, + 0.00527561130002141, + 0.0076497322879731655, + 0.010415669530630112, + 0.007083951495587826, + 0.007975513115525246, + 0.004385938867926598, + 0.01572766900062561, + 0.007787864655256271, + 0.00956498272716999, + 0.013528022915124893, + 0.00933835282921791, + 0.012580462731420994, + 0.013791725970804691, + 0.010580409318208694, + 0.016285894438624382, + 0.008896585553884506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.019487589597702026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025100174825638533 + ], + "opposite_logprob_tokens": [ + -6.002510070800781 + ], + "logprob_reward": -0.0025100174825638533, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025100174825638533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0683269202709198, + 0.2068738341331482, + 0.031487032771110535, + 0.166827991604805, + 0.03334195539355278, + 0.09312163293361664, + 0.18623298406600952, + 0.03756140545010567, + 0.15649911761283875, + 0.019727157428860664 + ], + "credit_map": [ + 0.011197198182344437, + 0.005730582866817713, + 0.011435720138251781, + 0.007176668848842382, + 0.005561642814427614, + 0.009613515809178352, + 0.007449401076883078, + 0.01016218401491642, + 0.030547166243195534, + 0.020127546042203903, + 0.019452383741736412, + 0.036766309291124344, + 0.022011300548911095, + 0.018704965710639954, + 0.022707847878336906, + 0.03655630722641945, + 0.004122389946132898, + 0.0033557305578142405, + 0.0029649296775460243, + 0.0030512288212776184, + 0.0036161066964268684, + 0.005964135285466909, + 0.004546954296529293, + 0.0038655558601021767, + 0.009129553101956844, + 0.02870725467801094, + 0.007694913074374199, + 0.01564643159508705, + 0.047134879976511, + 0.009741635993123055, + 0.021025624126195908, + 0.027747686952352524, + 0.0030781831592321396, + 0.004020394757390022, + 0.0027920883148908615, + 0.005020597018301487, + 0.004306970629841089, + 0.002501268405467272, + 0.007416892796754837, + 0.004205556586384773, + 0.004387564491480589, + 0.012851137667894363, + 0.003871534951031208, + 0.01121707446873188, + 0.011752113699913025, + 0.005680212751030922, + 0.032503608614206314, + 0.010858379304409027, + 0.005653987172991037, + 0.0782729759812355, + 0.0033774208277463913, + 0.005150098353624344, + 0.07507883757352829, + 0.0063090636394917965, + 0.007094035390764475, + 0.005296539980918169, + 0.0020394513849169016, + 0.005349546205252409, + 0.004182507749646902, + 0.0033551801461726427, + 0.007376781664788723, + 0.002971662674099207, + 0.006169166881591082, + 0.006117102690041065, + 0.013201585970818996, + 0.0204898901283741, + 0.016728181391954422, + 0.013488305732607841, + 0.016731616109609604, + 0.01452957559376955, + 0.019061069935560226, + 0.04226888343691826, + 0.0021301107481122017, + 0.003011078340932727, + 0.0018550589447841048, + 0.003172979922965169, + 0.0035160318948328495, + 0.001398377469740808, + 0.003156846622005105, + 0.0014866719720885158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.0025100174825638533 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054971784353256226 + ], + "opposite_logprob_tokens": [ + -2.929971694946289 + ], + "logprob_reward": -0.054971784353256226, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054971784353256226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20256809890270233, + 0.25523534417152405, + 0.08965650200843811, + 0.147531196475029, + 0.033121537417173386, + 0.047594103962183, + 0.05493571609258652, + 0.07328386604785919, + 0.06956060975790024, + 0.026513073593378067 + ], + "credit_map": [ + 0.029336437582969666, + 0.013686103746294975, + 0.04532276466488838, + 0.015503382310271263, + 0.01629807986319065, + 0.02817595563828945, + 0.022306222468614578, + 0.031939148902893066, + 0.04209579527378082, + 0.027751093730330467, + 0.02340819127857685, + 0.03210756927728653, + 0.028215399011969566, + 0.024892190471291542, + 0.025326713919639587, + 0.05143839865922928, + 0.022722313180565834, + 0.007868916727602482, + 0.006121750921010971, + 0.006656991317868233, + 0.0065063415095210075, + 0.01009424589574337, + 0.007523824460804462, + 0.022162118926644325, + 0.009592878632247448, + 0.026307307183742523, + 0.005811417009681463, + 0.01156528014689684, + 0.04389454796910286, + 0.007952185347676277, + 0.015371398068964481, + 0.02703617885708809, + 0.0034831815864890814, + 0.004795074462890625, + 0.0025753520894795656, + 0.004628676455467939, + 0.003896257607266307, + 0.0027278230991214514, + 0.006900942884385586, + 0.004114230163395405, + 0.0038112334441393614, + 0.008446857333183289, + 0.003402144182473421, + 0.0052467165514826775, + 0.008493172004818916, + 0.00408471142873168, + 0.008975584991276264, + 0.005133683327585459, + 0.0028830671217292547, + 0.009577884338796139, + 0.003621572395786643, + 0.005952363833785057, + 0.012261488474905491, + 0.0053458428010344505, + 0.009834454394876957, + 0.005459046456962824, + 0.004814109764993191, + 0.010776540264487267, + 0.009825876913964748, + 0.007286130916327238, + 0.013055609539151192, + 0.005202841479331255, + 0.012028799392282963, + 0.010293953120708466, + 0.006265469826757908, + 0.008768749423325062, + 0.007661300245672464, + 0.006522339303046465, + 0.007320186123251915, + 0.006236667279154062, + 0.009532174095511436, + 0.017253722995519638, + 0.002365807769820094, + 0.003188645699992776, + 0.002570482436567545, + 0.004177229944616556, + 0.0036780780646950006, + 0.0020119657274335623, + 0.005670518614351749, + 0.002850343007594347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.054971784353256226 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083347722887993 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011083347722887993, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083347722887993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046609971672296524, + 0.18327784538269043, + 0.0818798840045929, + 0.2728394567966461, + 0.025755152106285095, + 0.09159237891435623, + 0.08613418787717819, + 0.030873775482177734, + 0.11832025647163391, + 0.06271706521511078 + ], + "credit_map": [ + 0.008004892617464066, + 0.004064036533236504, + 0.0063793230801820755, + 0.004438030533492565, + 0.0042968736961483955, + 0.006556613836437464, + 0.005529304035007954, + 0.007340898271650076, + 0.03831951320171356, + 0.014582236297428608, + 0.014493530616164207, + 0.01921672187745571, + 0.01432784739881754, + 0.015110859647393227, + 0.01618374139070511, + 0.05104339122772217, + 0.023243041709065437, + 0.005386798642575741, + 0.006172984838485718, + 0.006799620110541582, + 0.007901692762970924, + 0.008034584112465382, + 0.009754577651619911, + 0.01458658091723919, + 0.016207827255129814, + 0.04487045481801033, + 0.011718260124325752, + 0.023422878235578537, + 0.07721760123968124, + 0.015087426640093327, + 0.03245304524898529, + 0.0518619641661644, + 0.0022811631206423044, + 0.002272551879286766, + 0.0018972065299749374, + 0.003513737814500928, + 0.003049221821129322, + 0.0030785866547375917, + 0.005124648567289114, + 0.00453803688287735, + 0.004478351213037968, + 0.012201107107102871, + 0.004292402416467667, + 0.011183671653270721, + 0.012987173162400723, + 0.0061506652273237705, + 0.03061319701373577, + 0.009685809724032879, + 0.004153480753302574, + 0.01945388875901699, + 0.005470572505146265, + 0.0075391619466245174, + 0.018954958766698837, + 0.005049137398600578, + 0.01817428693175316, + 0.007338705472648144, + 0.0018987200455740094, + 0.004657126497477293, + 0.004916129633784294, + 0.003469685325399041, + 0.005167522467672825, + 0.0021947792265564203, + 0.005261573474854231, + 0.0033082379959523678, + 0.010630673728883266, + 0.0149972103536129, + 0.013278291560709476, + 0.011111709289252758, + 0.012081803753972054, + 0.011023608967661858, + 0.014897681772708893, + 0.03029927983880043, + 0.005811367649585009, + 0.008937370963394642, + 0.006739698816090822, + 0.009468024596571922, + 0.010360471904277802, + 0.006337556056678295, + 0.009429318830370903, + 0.00563325360417366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.011083347722887993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42896732687950134 + ], + "opposite_logprob_tokens": [ + -1.0539673566818237 + ], + "logprob_reward": -0.42896732687950134, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42896732687950134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07451831549406052, + 0.14668740332126617, + 0.06184336543083191, + 0.23310622572898865, + 0.03517434746026993, + 0.05190526321530342, + 0.22537685930728912, + 0.03441068157553673, + 0.09575914591550827, + 0.04121837019920349 + ], + "credit_map": [ + 0.01538127101957798, + 0.004291782155632973, + 0.012563752941787243, + 0.006950838025659323, + 0.00505297863855958, + 0.01025552861392498, + 0.007437235210090876, + 0.012584930285811424, + 0.018093612045049667, + 0.015595514327287674, + 0.019215397536754608, + 0.021066199988126755, + 0.01799449324607849, + 0.018086951225996017, + 0.017367379739880562, + 0.01926785707473755, + 0.008938363753259182, + 0.008020767010748386, + 0.0038214235100895166, + 0.005634082015603781, + 0.008157363161444664, + 0.007297658361494541, + 0.007555550429970026, + 0.012418162077665329, + 0.013476168736815453, + 0.04519272595643997, + 0.010489285923540592, + 0.01906399056315422, + 0.06889110058546066, + 0.013111847452819347, + 0.02611609175801277, + 0.036765024065971375, + 0.0026710943784564734, + 0.0037432173267006874, + 0.002719657029956579, + 0.0051587470807135105, + 0.004124379251152277, + 0.0039803762920200825, + 0.007411502301692963, + 0.0053653777576982975, + 0.0026687942445278168, + 0.009694812819361687, + 0.0032579272519797087, + 0.005344374570995569, + 0.0071869813837111, + 0.0033679308835417032, + 0.013710267841815948, + 0.006674179341644049, + 0.007501176092773676, + 0.09018413722515106, + 0.005747199058532715, + 0.0072577837854623795, + 0.08602485060691833, + 0.007664868142455816, + 0.013528679497539997, + 0.007468174211680889, + 0.0026216753758490086, + 0.00600105244666338, + 0.004442998208105564, + 0.0036914281081408262, + 0.006379186175763607, + 0.002250485122203827, + 0.005007638130337, + 0.004016217775642872, + 0.009330565109848976, + 0.011400128714740276, + 0.012270585633814335, + 0.008211673237383366, + 0.00910524558275938, + 0.010089261457324028, + 0.010229070670902729, + 0.025122622027993202, + 0.004260418005287647, + 0.005740192718803883, + 0.004690160043537617, + 0.0054078777320683, + 0.006765349768102169, + 0.0026379625778645277, + 0.007483040913939476, + 0.0042333705350756645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.42896732687950134 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, top-down macro shot remains static, focused on a perfectly square block of cold, pale yellow butter in the center of a dark, matte cast iron skillet. Under the influence of intense, invisible heat from below, the butter's sharp edges instantly begin to soften and shimmer. A ring of transparent, shimmering liquid forms at its base, rapidly expanding as the solid structure inexorably collapses and collapses inward. The transformation accelerates, and the entire pie dissolves into a vibrant pool of seething, translucent golden liquid. Tiny, rippled bubbles erupt across the surface in a chaotic dance, creating a dynamic, fizzing texture that catches the bright, direct overhead light, giving the entire pool a hypnotic glow. The scene captures the swift and dramatic metamorphosis, creating a concentrated, intensely sensory atmosphere of pure thermal energy and culinary action.", + "seed": 10601, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_6/index_9052_seed_10601_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.46990127945318816, + "vlm_logprob": -0.46990127945318816, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.46990127945318816, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18238656222820282 + ], + "opposite_logprob_tokens": [ + -1.9323865175247192 + ], + "logprob_reward": -0.18238656222820282, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.18238656222820282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22588129341602325, + 0.07467921823263168, + 0.17076866328716278, + 0.0684180036187172, + 0.11551228910684586, + 0.07839798927307129, + 0.12675787508487701, + 0.049472857266664505, + 0.04049426317214966, + 0.049617499113082886 + ], + "credit_map": [ + 0.04843679070472717, + 0.014095810241997242, + 0.03478012606501579, + 0.01673826202750206, + 0.01417198870331049, + 0.019067035987973213, + 0.027950318530201912, + 0.05064094439148903, + 0.00957401655614376, + 0.005838139448314905, + 0.012581187300384045, + 0.010269821621477604, + 0.006255395244807005, + 0.005950354505330324, + 0.012639029882848263, + 0.011571269482374191, + 0.022383805364370346, + 0.010906592942774296, + 0.04940500110387802, + 0.018697163090109825, + 0.015488819219172001, + 0.01220699306577444, + 0.014731645584106445, + 0.02694864384829998, + 0.006747846025973558, + 0.008382012136280537, + 0.0081639951094985, + 0.008754560723900795, + 0.009423676878213882, + 0.008669199422001839, + 0.01050624344497919, + 0.007770468946546316, + 0.018448980525135994, + 0.008837551809847355, + 0.013518458232283592, + 0.011213123798370361, + 0.01277568843215704, + 0.0140919154509902, + 0.01317150890827179, + 0.023455068469047546, + 0.012720868922770023, + 0.007192081771790981, + 0.010432802140712738, + 0.01562274806201458, + 0.006712150759994984, + 0.007788892835378647, + 0.009718593209981918, + 0.008209855295717716, + 0.017664914950728416, + 0.014770963229238987, + 0.01826454885303974, + 0.016252629458904266, + 0.01236000470817089, + 0.018471498042345047, + 0.009461820125579834, + 0.019511498510837555, + 0.007496450562030077, + 0.00746338302269578, + 0.006877755280584097, + 0.006323761772364378, + 0.005676125641912222, + 0.007579177152365446, + 0.0039760032668709755, + 0.004080198705196381, + 0.004927231464534998, + 0.005218187812715769, + 0.004553903825581074, + 0.006694422569125891, + 0.005178617779165506, + 0.00525868171826005, + 0.00547333387658, + 0.0031898850575089455, + 0.007521134801208973, + 0.006124631501734257, + 0.006110413931310177, + 0.0058541023172438145, + 0.006415620446205139, + 0.004875992424786091, + 0.008294074796140194, + 0.004421525169163942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.18238656222820282 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046732728369534016 + ], + "opposite_logprob_tokens": [ + -5.379673480987549 + ], + "logprob_reward": -0.0046732728369534016, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046732728369534016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13748247921466827, + 0.06111226975917816, + 0.04417385905981064, + 0.10632313787937164, + 0.07808806002140045, + 0.09114431589841843, + 0.2862548530101776, + 0.055933352559804916, + 0.04332394525408745, + 0.09616374969482422 + ], + "credit_map": [ + 0.021247198805212975, + 0.010726598091423512, + 0.021487390622496605, + 0.019241327419877052, + 0.014166846871376038, + 0.01454851869493723, + 0.016776220872998238, + 0.01928836479783058, + 0.00833720900118351, + 0.004277244210243225, + 0.009652474895119667, + 0.010066994465887547, + 0.00452836649492383, + 0.006213461048901081, + 0.006803978234529495, + 0.011232532560825348, + 0.004994458518922329, + 0.0060303835198283195, + 0.007944010198116302, + 0.004020572640001774, + 0.008181017823517323, + 0.004194420762360096, + 0.003851413493975997, + 0.0049575804732739925, + 0.010249766521155834, + 0.011246324516832829, + 0.012121560983359814, + 0.013897174969315529, + 0.01650717668235302, + 0.012063081376254559, + 0.01659001223742962, + 0.01364803221076727, + 0.00854252278804779, + 0.008420164696872234, + 0.009569021873176098, + 0.009068749845027924, + 0.01045195572078228, + 0.011687652207911015, + 0.010414048098027706, + 0.009933938272297382, + 0.013981442898511887, + 0.01046849973499775, + 0.011232564225792885, + 0.01593579538166523, + 0.008765513077378273, + 0.008794884197413921, + 0.011867905035614967, + 0.010097699239850044, + 0.035476185381412506, + 0.03518948704004288, + 0.043475087732076645, + 0.04216398671269417, + 0.02518162690103054, + 0.043860629200935364, + 0.0257964376360178, + 0.03511138632893562, + 0.00591211998835206, + 0.009031839668750763, + 0.005486922804266214, + 0.005760345607995987, + 0.006367966532707214, + 0.010043255053460598, + 0.007749001495540142, + 0.005581896286457777, + 0.005095835775136948, + 0.0053359768353402615, + 0.0057925512082874775, + 0.005552449729293585, + 0.004998527932912111, + 0.005245069041848183, + 0.0057921805419027805, + 0.005511350464075804, + 0.01129528321325779, + 0.014376973733305931, + 0.01128165703266859, + 0.00951431505382061, + 0.010987850837409496, + 0.01514661218971014, + 0.011773956939578056, + 0.01178708765655756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.0046732728369534016 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026369068771600723 + ], + "opposite_logprob_tokens": [ + -3.651369094848633 + ], + "logprob_reward": -0.026369068771600723, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026369068771600723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16815562546253204, + 0.08658484369516373, + 0.12308719009160995, + 0.042060595005750656, + 0.039032433182001114, + 0.1168341189622879, + 0.3120315670967102, + 0.04956258833408356, + 0.03111100010573864, + 0.03154008835554123 + ], + "credit_map": [ + 0.020150698721408844, + 0.00950388889759779, + 0.029015759006142616, + 0.029716534540057182, + 0.010124788619577885, + 0.010825131088495255, + 0.021036939695477486, + 0.03778187185525894, + 0.011879554018378258, + 0.006298450753092766, + 0.014234443195164204, + 0.015185347758233547, + 0.006681900005787611, + 0.006694591138511896, + 0.012217115610837936, + 0.013393435627222061, + 0.013164494186639786, + 0.012747262604534626, + 0.028958138078451157, + 0.01285175234079361, + 0.017862405627965927, + 0.009249350987374783, + 0.009042179211974144, + 0.019211601465940475, + 0.005511222407221794, + 0.003935441840440035, + 0.00570580642670393, + 0.004711391869932413, + 0.005948770325630903, + 0.005225518252700567, + 0.004458911716938019, + 0.006563530769199133, + 0.005648977588862181, + 0.004753813613206148, + 0.004725615959614515, + 0.004081322345882654, + 0.004277616273611784, + 0.0055677806958556175, + 0.004421844147145748, + 0.005555460229516029, + 0.01974574103951454, + 0.014624658972024918, + 0.012556513771414757, + 0.01785845123231411, + 0.011008675210177898, + 0.013805177994072437, + 0.014140668325126171, + 0.013094224035739899, + 0.04807737097144127, + 0.04313252121210098, + 0.03830844536423683, + 0.039335496723651886, + 0.027723699808120728, + 0.05428418889641762, + 0.02643590420484543, + 0.03473392874002457, + 0.006258280482143164, + 0.0070994761772453785, + 0.007409753743559122, + 0.008344150148332119, + 0.006200585514307022, + 0.0048682671040296555, + 0.004854731727391481, + 0.004527341574430466, + 0.004450429696589708, + 0.0033107115887105465, + 0.005673680920153856, + 0.006049139425158501, + 0.003513477509841323, + 0.0019449908286333084, + 0.0026757235173135996, + 0.003492843359708786, + 0.004316913895308971, + 0.0037958440370857716, + 0.00403338298201561, + 0.00522797042503953, + 0.0036996430717408657, + 0.0034052436240017414, + 0.0034827953204512596, + 0.0035782891791313887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.026369068771600723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103772640228271 + ], + "opposite_logprob_tokens": [ + -0.16037724912166595 + ], + "logprob_reward": -1.9103772640228271, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103772640228271, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17893801629543304, + 0.048041973263025284, + 0.10879058390855789, + 0.044254120439291, + 0.1282730996608734, + 0.10286074876785278, + 0.23563529551029205, + 0.0419127494096756, + 0.027096783742308617, + 0.08419665694236755 + ], + "credit_map": [ + 0.030108390375971794, + 0.013658326119184494, + 0.025268636643886566, + 0.0187296774238348, + 0.018197454512119293, + 0.014921172522008419, + 0.030140338465571404, + 0.02791401371359825, + 0.0070909736678004265, + 0.0037268472369760275, + 0.006734916009008884, + 0.0054034083150327206, + 0.003691696561872959, + 0.00563674746081233, + 0.007303280755877495, + 0.008454104885458946, + 0.011273836717009544, + 0.00784666184335947, + 0.021215299144387245, + 0.014214331284165382, + 0.01007574237883091, + 0.010049030184745789, + 0.0170404352247715, + 0.017075246199965477, + 0.005130676552653313, + 0.005163556430488825, + 0.006144378334283829, + 0.004532600753009319, + 0.005683934781700373, + 0.0055649131536483765, + 0.0054460070095956326, + 0.00658805388957262, + 0.016045087948441505, + 0.013599036261439323, + 0.015152497217059135, + 0.01311268750578165, + 0.015052901580929756, + 0.017792725935578346, + 0.0187266543507576, + 0.018791504204273224, + 0.01940115913748741, + 0.010692354291677475, + 0.012450472451746464, + 0.015241793356835842, + 0.01041968259960413, + 0.010193584486842155, + 0.010611475445330143, + 0.013850224204361439, + 0.029031014069914818, + 0.02723773568868637, + 0.03653746843338013, + 0.03701018542051315, + 0.020650716498494148, + 0.034429341554641724, + 0.02117891050875187, + 0.029559915885329247, + 0.005406189244240522, + 0.006631409749388695, + 0.004585073795169592, + 0.005817585624754429, + 0.004672737792134285, + 0.00638334546238184, + 0.004673854447901249, + 0.003742554225027561, + 0.0030947525519877672, + 0.0045640235766768456, + 0.0026831883005797863, + 0.0028643591795116663, + 0.003327520564198494, + 0.004199035931378603, + 0.0033251435961574316, + 0.003038761205971241, + 0.009110309183597565, + 0.014326250180602074, + 0.008059150539338589, + 0.009192549623548985, + 0.011085222475230694, + 0.010412996634840965, + 0.01308891549706459, + 0.008921264670789242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -1.9103772640228271 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up, top-down macro view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a perfectly square block of pale yellow butter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the butter placed in a dark, matte cast iron skillet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butter rapidly melt into a pool of golden liquid with rippled bubbles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera static throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2257002294063568 + ], + "opposite_logprob_tokens": [ + -1.6007002592086792 + ], + "logprob_reward": -0.2257002294063568, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2257002294063568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12882190942764282, + 0.059451278299093246, + 0.11163876205682755, + 0.038704920560121536, + 0.1472834348678589, + 0.04846236854791641, + 0.39050838351249695, + 0.02939346432685852, + 0.01865818351507187, + 0.027077272534370422 + ], + "credit_map": [ + 0.024102002382278442, + 0.008914045989513397, + 0.017871584743261337, + 0.015739474445581436, + 0.011659684590995312, + 0.011116057634353638, + 0.01728721708059311, + 0.022131849080324173, + 0.010110614821314812, + 0.004464638419449329, + 0.007395156659185886, + 0.006208107341080904, + 0.00509247463196516, + 0.00686099287122488, + 0.0077113136649131775, + 0.011607982218265533, + 0.01409953273832798, + 0.006737599149346352, + 0.03200220316648483, + 0.013738006353378296, + 0.011130983941257, + 0.007122762035578489, + 0.010394985787570477, + 0.016412686556577682, + 0.004121390171349049, + 0.004375775810331106, + 0.005379976239055395, + 0.003266931977123022, + 0.004854657221585512, + 0.007449333090335131, + 0.003974552731961012, + 0.005282304249703884, + 0.017872709780931473, + 0.014644404873251915, + 0.01792321726679802, + 0.017048103734850883, + 0.015453810803592205, + 0.02016862854361534, + 0.022654321044683456, + 0.021518239751458168, + 0.008233604952692986, + 0.008370181545615196, + 0.005089264828711748, + 0.005544126033782959, + 0.007216943893581629, + 0.0046556550078094006, + 0.004545891657471657, + 0.004806701559573412, + 0.05904147028923035, + 0.05573837831616402, + 0.04507298022508621, + 0.04778456687927246, + 0.03545139357447624, + 0.0713891014456749, + 0.03259294852614403, + 0.04343755915760994, + 0.005577140487730503, + 0.0028016187716275454, + 0.004851652774959803, + 0.003763164859265089, + 0.0027708096895366907, + 0.003520226338878274, + 0.0027572091203182936, + 0.0033516420517116785, + 0.0029027468990534544, + 0.00233355350792408, + 0.0022317105904221535, + 0.003363729687407613, + 0.0021539742592722178, + 0.0021997548174113035, + 0.0019009100506082177, + 0.0015718038193881512, + 0.0025921352207660675, + 0.004175733774900436, + 0.0021853537764400244, + 0.0034705267753452063, + 0.0036348639987409115, + 0.0033595904242247343, + 0.004210519138723612, + 0.003448548261076212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.2257002294063568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_tensor([9334, 9334])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_tensor([9334, 9334])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0d219eeaaaeceabac0e6f3536f29beff61a10a21 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_tensor([9334, 9334])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme macro close-up, the camera remains still, its focus locked precisely on a new fern leaf as it slowly unfurls from its signature curl. A vibrant emerald green, covered in a dense layer of fine white hairs, the leaf appears vibrant in the soft, mist-diffused golden light of dawn. Precariously suspended at the tip of the leaf, a perfectly spherical dewdrop shimmers, its crystal-clear form like a miniature crystal, its limpid interior reflecting and distorting the out-of-focus forest backdrop, shimmering with light. As a barely perceptible breeze sweeps across the leaf, a subtle yet constant tremor ripples through the leaf, a movement so slight that it sways the dewdrop itself, causing light to flow steadily and deliberately across its surface. A thick, translucent morning mist gently swirls in the background, obscuring everything beyond, leaving only the soft hues of moss and damp soil, creating an atmosphere of profound tranquility and new hope.", + "seed": 26017, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_9334_seed_26017_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09352658041752875, + "vlm_logprob": -0.09352658041752875, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.09352658041752875, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020982317626476288 + ], + "opposite_logprob_tokens": [ + -5.020982265472412 + ], + "logprob_reward": -0.020982317626476288, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.020982317626476288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21690648794174194, + 0.16085684299468994, + 0.17479950189590454, + 0.08556875586509705, + 0.06395194679498672, + 0.04488098993897438, + 0.0723765417933464, + 0.037958234548568726, + 0.09503421932458878, + 0.047666534781455994 + ], + "credit_map": [ + 0.0205905232578516, + 0.026361815631389618, + 0.023261044174432755, + 0.031164245679974556, + 0.035813190042972565, + 0.022224871441721916, + 0.02971787005662918, + 0.027772922068834305, + 0.016307508572936058, + 0.018367821350693703, + 0.016595426946878433, + 0.024127328768372536, + 0.03150487318634987, + 0.020030552521348, + 0.01803562231361866, + 0.015887713059782982, + 0.015340288169682026, + 0.02166730910539627, + 0.024291502311825752, + 0.023162981495261192, + 0.024278441444039345, + 0.012578552588820457, + 0.038129743188619614, + 0.015350691974163055, + 0.005813174415379763, + 0.00797285046428442, + 0.010388541035354137, + 0.014106325805187225, + 0.01307407021522522, + 0.011135225184261799, + 0.013806966133415699, + 0.009271600283682346, + 0.005025489255785942, + 0.0071916719898581505, + 0.008936635218560696, + 0.00784971658140421, + 0.009502368979156017, + 0.009588483721017838, + 0.008264363743364811, + 0.007593216374516487, + 0.0037194194737821817, + 0.005287667270749807, + 0.005937487818300724, + 0.007088087499141693, + 0.0053626904264092445, + 0.004383701365441084, + 0.00810542143881321, + 0.004996513947844505, + 0.006950910668820143, + 0.009169734083116055, + 0.008357962593436241, + 0.011298509314656258, + 0.009533140808343887, + 0.005982424132525921, + 0.015112065710127354, + 0.005971796810626984, + 0.002997918287292123, + 0.004426742438226938, + 0.004613572731614113, + 0.006612695753574371, + 0.004159025382250547, + 0.00437550712376833, + 0.006432127207517624, + 0.004340645857155323, + 0.007557297591120005, + 0.013245375826954842, + 0.009728776291012764, + 0.017129523679614067, + 0.010752328671514988, + 0.010876295156776905, + 0.014794331043958664, + 0.010950297117233276, + 0.0035609868355095387, + 0.005292558576911688, + 0.005188231356441975, + 0.006206748075783253, + 0.00710734399035573, + 0.006826332770287991, + 0.0071570491418242455, + 0.006327283568680286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.020982317626476288 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906280249357224 + ], + "opposite_logprob_tokens": [ + -2.4640626907348633 + ], + "logprob_reward": -0.08906280249357224, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906280249357224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09635210037231445, + 0.5367336273193359, + 0.025300953537225723, + 0.023336080834269524, + 0.04410409927368164, + 0.09487694501876831, + 0.06249111145734787, + 0.0509486049413681, + 0.05352577939629555, + 0.012330710887908936 + ], + "credit_map": [ + 0.006796440575271845, + 0.009324322454631329, + 0.008885045535862446, + 0.017079360783100128, + 0.01383338775485754, + 0.013299981132149696, + 0.011019635945558548, + 0.016113925725221634, + 0.043875351548194885, + 0.05277881398797035, + 0.05871818959712982, + 0.055571410804986954, + 0.1274786740541458, + 0.07802979648113251, + 0.06397183239459991, + 0.05630955472588539, + 0.0018234552117064595, + 0.0024738560896366835, + 0.0030895499512553215, + 0.0028040558099746704, + 0.0052924673072993755, + 0.003401912981644273, + 0.00335497222840786, + 0.0030606843065470457, + 0.0020952087361365557, + 0.003085631877183914, + 0.003348143771290779, + 0.0030031204223632812, + 0.0024802640546113253, + 0.0029527046717703342, + 0.0030778544023633003, + 0.0032931522000581026, + 0.00347305228933692, + 0.00515148788690567, + 0.004670407623052597, + 0.007602551020681858, + 0.006508734077215195, + 0.004572203382849693, + 0.007325117010623217, + 0.004800545051693916, + 0.009160005487501621, + 0.011173168197274208, + 0.01319358590990305, + 0.011997337453067303, + 0.013263567350804806, + 0.008306519128382206, + 0.01850399188697338, + 0.009278764948248863, + 0.004571785219013691, + 0.007388219237327576, + 0.007263466250151396, + 0.008761769160628319, + 0.008234505541622639, + 0.006350942887365818, + 0.014237905852496624, + 0.0056825196370482445, + 0.003330652602016926, + 0.004694389645010233, + 0.006308138836175203, + 0.0047713578678667545, + 0.0078580342233181, + 0.004961678758263588, + 0.0138844670727849, + 0.0051398868672549725, + 0.0029632204677909613, + 0.0052766804583370686, + 0.006168443709611893, + 0.0062016588635742664, + 0.006626185029745102, + 0.004519727546721697, + 0.014834985136985779, + 0.00693487748503685, + 0.0008768501575104892, + 0.0016205046558752656, + 0.0015281285159289837, + 0.0019758979324251413, + 0.001252787304110825, + 0.0013282798463478684, + 0.002266881987452507, + 0.001481379964388907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.08906280249357224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043076761066913605 + ], + "opposite_logprob_tokens": [ + -3.168076753616333 + ], + "logprob_reward": -0.043076761066913605, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043076761066913605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2714158892631531, + 0.18214891850948334, + 0.11345276981592178, + 0.09576275199651718, + 0.07065289467573166, + 0.05983654037117958, + 0.08925602585077286, + 0.049622900784015656, + 0.05116264149546623, + 0.016688624396920204 + ], + "credit_map": [ + 0.01617373898625374, + 0.028245829045772552, + 0.03140044957399368, + 0.042718950659036636, + 0.046825092285871506, + 0.02844930998980999, + 0.045721638947725296, + 0.03188089653849602, + 0.019431889057159424, + 0.019462229683995247, + 0.022496527060866356, + 0.018872126936912537, + 0.03582463413476944, + 0.023021696135401726, + 0.018902137875556946, + 0.02413768507540226, + 0.008472704328596592, + 0.016446320340037346, + 0.012640325352549553, + 0.016449429094791412, + 0.014371441677212715, + 0.010211003012955189, + 0.023294227197766304, + 0.011567320674657822, + 0.007998868823051453, + 0.010296971537172794, + 0.014752398245036602, + 0.00960483681410551, + 0.015159728936851025, + 0.010149454697966576, + 0.01798432320356369, + 0.00981617346405983, + 0.005762024316936731, + 0.008728742599487305, + 0.007949255406856537, + 0.00877432432025671, + 0.008950658142566681, + 0.005063366610556841, + 0.017710547894239426, + 0.007713975850492716, + 0.0050711012445390224, + 0.0065962811931967735, + 0.0088191581889987, + 0.005956323817372322, + 0.008689407259225845, + 0.004692358430474997, + 0.013963363133370876, + 0.006048545241355896, + 0.007845627143979073, + 0.010028215125203133, + 0.010515328496694565, + 0.01115078292787075, + 0.012427707202732563, + 0.009805719368159771, + 0.018763946369290352, + 0.008718697354197502, + 0.002874664030969143, + 0.0055318558588624, + 0.00557232229039073, + 0.0055493442341685295, + 0.006886213086545467, + 0.006156101357191801, + 0.011965219862759113, + 0.005087177734822035, + 0.00349582196213305, + 0.0060547515749931335, + 0.005880720913410187, + 0.007534153293818235, + 0.006688916590064764, + 0.004499920643866062, + 0.010462447069585323, + 0.006545906886458397, + 0.0011068243766203523, + 0.0020672809332609177, + 0.002153577283024788, + 0.0020998597610741854, + 0.0019641888793557882, + 0.0017297844169661403, + 0.004117690492421389, + 0.0014494182541966438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.043076761066913605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010452766437083483 + ], + "opposite_logprob_tokens": [ + -6.876045227050781 + ], + "logprob_reward": -0.0010452766437083483, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010452766437083483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1754635125398636, + 0.11147857457399368, + 0.17507198452949524, + 0.0998765304684639, + 0.053084515035152435, + 0.07519035786390305, + 0.176510289311409, + 0.06565357744693756, + 0.04364810138940811, + 0.024022527039051056 + ], + "credit_map": [ + 0.014304263517260551, + 0.01872766762971878, + 0.022734399884939194, + 0.030424239113926888, + 0.022394996136426926, + 0.01341486256569624, + 0.028246430680155754, + 0.02521664835512638, + 0.009117169305682182, + 0.010904397815465927, + 0.012122506275773048, + 0.013968339189887047, + 0.017927147448062897, + 0.01681031845510006, + 0.011330204084515572, + 0.019298497587442398, + 0.011620333418250084, + 0.022683948278427124, + 0.02220865525305271, + 0.02780403196811676, + 0.024710986763238907, + 0.013292595744132996, + 0.03833101689815521, + 0.014420409686863422, + 0.006407782435417175, + 0.009255158714950085, + 0.011358469724655151, + 0.01575174182653427, + 0.01713648997247219, + 0.016433358192443848, + 0.01130184717476368, + 0.012231683358550072, + 0.003314775414764881, + 0.003818634431809187, + 0.006789235398173332, + 0.00861311424523592, + 0.009439419955015182, + 0.005749016534537077, + 0.009101856499910355, + 0.006258463487029076, + 0.005793612450361252, + 0.00842301081866026, + 0.008690300397574902, + 0.01081018429249525, + 0.009976583532989025, + 0.007171823643147945, + 0.015191296115517616, + 0.009133552201092243, + 0.014722492545843124, + 0.021404776722192764, + 0.021310845389962196, + 0.022073334082961082, + 0.02560405060648918, + 0.014966248534619808, + 0.04120732843875885, + 0.015221210196614265, + 0.004176292568445206, + 0.008772152476012707, + 0.008098440244793892, + 0.0086784977465868, + 0.008890806697309017, + 0.011249840259552002, + 0.009100709110498428, + 0.006686839275062084, + 0.0029685888439416885, + 0.004187700804322958, + 0.005466016475111246, + 0.007565754931420088, + 0.005937556736171246, + 0.005252845585346222, + 0.006015389692038298, + 0.006254248321056366, + 0.0022876255679875612, + 0.00296505494043231, + 0.0033306193072348833, + 0.0026411053258925676, + 0.0026000840589404106, + 0.002645886270329356, + 0.004465979989618063, + 0.0030861718114465475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.0010452766437083483 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31346574425697327 + ], + "opposite_logprob_tokens": [ + -1.3134657144546509 + ], + "logprob_reward": -0.31346574425697327, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31346574425697327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1129683330655098, + 0.17355266213417053, + 0.21515020728111267, + 0.07788845151662827, + 0.05320662260055542, + 0.1998567283153534, + 0.04790722578763962, + 0.046912066638469696, + 0.02763940393924713, + 0.044918276369571686 + ], + "credit_map": [ + 0.01110823079943657, + 0.016914179548621178, + 0.011114885099232197, + 0.020416676998138428, + 0.014334090985357761, + 0.011063551530241966, + 0.011253499425947666, + 0.016763223335146904, + 0.014058321714401245, + 0.016579490154981613, + 0.016179531812667847, + 0.02441793866455555, + 0.03053831309080124, + 0.0234982892870903, + 0.022865628823637962, + 0.025415144860744476, + 0.017530567944049835, + 0.023405225947499275, + 0.02945978008210659, + 0.03403367102146149, + 0.028771191835403442, + 0.019560163840651512, + 0.04184480383992195, + 0.020544802770018578, + 0.0051187193021178246, + 0.0070349229499697685, + 0.01005947683006525, + 0.010364839807152748, + 0.014615245163440704, + 0.010103380307555199, + 0.010674173943698406, + 0.009917699731886387, + 0.004083278588950634, + 0.008230315521359444, + 0.006028616800904274, + 0.007676920387893915, + 0.0066481782123446465, + 0.006245747674256563, + 0.00645972928032279, + 0.007833833806216717, + 0.019176926463842392, + 0.02137455716729164, + 0.0242606308311224, + 0.019284136593341827, + 0.03321138024330139, + 0.025934619829058647, + 0.034936126321554184, + 0.021678343415260315, + 0.004180597607046366, + 0.005976856220513582, + 0.005845780950039625, + 0.0056396001018583775, + 0.006550392135977745, + 0.005535309202969074, + 0.010156845673918724, + 0.004021843895316124, + 0.003731998149305582, + 0.004829822573810816, + 0.005433098413050175, + 0.0055933562107384205, + 0.006861168425530195, + 0.006507402285933495, + 0.008806241676211357, + 0.005148977506905794, + 0.0018597595626488328, + 0.003239601617679, + 0.003399315057322383, + 0.004869750700891018, + 0.0035694993566721678, + 0.0028817313723266125, + 0.005176002625375986, + 0.002643743297085166, + 0.0029385974630713463, + 0.0047326842322945595, + 0.0053321681916713715, + 0.0063552698120474815, + 0.005102813709527254, + 0.0034373863600194454, + 0.011647204868495464, + 0.005372151266783476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.31346574425697327 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up, the camera remains still, its focus locked precisely on a new fern leaf as it slowly unfurls from its signature curl. A vibrant emerald green, covered in a dense layer of fine white hairs, the leaf appears vibrant in the soft, mist-diffused golden light of dawn. Precariously suspended at the tip of the leaf, a perfectly spherical dewdrop shimmers, its crystal-clear form like a miniature crystal, its limpid interior reflecting and distorting the out-of-focus forest backdrop, shimmering with light. As a barely perceptible breeze sweeps across the leaf, a subtle yet constant tremor ripples through the leaf, a movement so slight that it sways the dewdrop itself, causing light to flow steadily and deliberately across its surface. A thick, translucent morning mist gently swirls in the background, obscuring everything beyond, leaving only the soft hues of moss and damp soil, creating an atmosphere of profound tranquility and new hope.", + "seed": 26018, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_9334_seed_26018_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.46283425986766813, + "vlm_logprob": -0.46283425986766813, + "vlm_yes_no_margin": 1.575, + "vlm_token_credit": -0.46283425986766813, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05604549124836922 + ], + "opposite_logprob_tokens": [ + -3.5560455322265625 + ], + "logprob_reward": -0.05604549124836922, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.05604549124836922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10565511882305145, + 0.019008364528417587, + 0.026966610923409462, + 0.02615405060350895, + 0.044764574617147446, + 0.04458186402916908, + 0.09095393866300583, + 0.31215158104896545, + 0.28539690375328064, + 0.04436697065830231 + ], + "credit_map": [ + 0.014909620396792889, + 0.007481980137526989, + 0.022409092634916306, + 0.008720938116312027, + 0.009518272243440151, + 0.02070944383740425, + 0.010623708367347717, + 0.011282059364020824, + 0.002286230679601431, + 0.001710011507384479, + 0.00271420506760478, + 0.002661576261743903, + 0.0020941223483532667, + 0.0026201612781733274, + 0.002321915701031685, + 0.002600142266601324, + 0.004239425994455814, + 0.001888542901724577, + 0.003993875812739134, + 0.002624321263283491, + 0.0024995505809783936, + 0.006140247453004122, + 0.0025135097093880177, + 0.00306714023463428, + 0.0030721821822226048, + 0.0023113773204386234, + 0.004355901386588812, + 0.003208445617929101, + 0.0035461036022752523, + 0.003104808507487178, + 0.0034570780117064714, + 0.003098154440522194, + 0.004950305446982384, + 0.0032840322237461805, + 0.011297866702079773, + 0.004945698194205761, + 0.004391972906887531, + 0.007522928528487682, + 0.0038399496115744114, + 0.004531823564320803, + 0.00434027099981904, + 0.00591569347307086, + 0.00630974443629384, + 0.008176267147064209, + 0.006077635567635298, + 0.004500206094235182, + 0.004862388130277395, + 0.00439965445548296, + 0.013838192448019981, + 0.008109004236757755, + 0.016807088628411293, + 0.010303289629518986, + 0.010561774484813213, + 0.013123270124197006, + 0.010019606910645962, + 0.008191713131964207, + 0.10739089548587799, + 0.012432879768311977, + 0.02029297687113285, + 0.01476781815290451, + 0.01782631129026413, + 0.10718511790037155, + 0.011761761270463467, + 0.02049381472170353, + 0.06471918523311615, + 0.027806082740426064, + 0.036674853414297104, + 0.05150958150625229, + 0.02574615180492401, + 0.040841907262802124, + 0.01899830438196659, + 0.019100820645689964, + 0.007491094991564751, + 0.0047484091483056545, + 0.005078580696135759, + 0.00968236941844225, + 0.003545569721609354, + 0.004888902883976698, + 0.004129126667976379, + 0.00480292085558176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.05604549124836922 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6006124019622803 + ], + "opposite_logprob_tokens": [ + -0.22561243176460266 + ], + "logprob_reward": -1.6006124019622803, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6006124019622803, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2635051906108856, + 0.07480013370513916, + 0.046637292951345444, + 0.050810180604457855, + 0.03582210838794708, + 0.0404091440141201, + 0.03560483828186989, + 0.09808485209941864, + 0.2945268154144287, + 0.05979936197400093 + ], + "credit_map": [ + 0.029632875695824623, + 0.023327432572841644, + 0.03385288640856743, + 0.029398873448371887, + 0.03844268620014191, + 0.03889400511980057, + 0.03942113369703293, + 0.030535336583852768, + 0.010646301321685314, + 0.005701110232621431, + 0.008590259589254856, + 0.008294514380395412, + 0.008934400975704193, + 0.016497593373060226, + 0.008633089251816273, + 0.007502880413085222, + 0.006808421574532986, + 0.0032940064556896687, + 0.006441215984523296, + 0.004008356481790543, + 0.0052999104373157024, + 0.010884848423302174, + 0.0052824467420578, + 0.004618093371391296, + 0.005281954538077116, + 0.005050170700997114, + 0.005856087431311607, + 0.006966109853237867, + 0.006277901586145163, + 0.009079447016119957, + 0.005978994537144899, + 0.006319522857666016, + 0.0028799010906368494, + 0.0035724907647818327, + 0.004882434383034706, + 0.0050392779521644115, + 0.0050574676133692265, + 0.0039909640327095985, + 0.004505696706473827, + 0.005893881432712078, + 0.004134293645620346, + 0.0037236178759485483, + 0.004369273316115141, + 0.005949138663709164, + 0.006017704028636217, + 0.004455422051250935, + 0.005699691828340292, + 0.006060009356588125, + 0.0037607743870466948, + 0.003022483317181468, + 0.0042594922706484795, + 0.004472727421671152, + 0.005524191539734602, + 0.004478673916310072, + 0.0040354193188250065, + 0.006051081698387861, + 0.01970292069017887, + 0.010739191435277462, + 0.011528722941875458, + 0.014195147901773453, + 0.008705440908670425, + 0.01859319396317005, + 0.007089770864695311, + 0.007530475966632366, + 0.05772542953491211, + 0.02809966914355755, + 0.04379512742161751, + 0.055374860763549805, + 0.023719536140561104, + 0.0350516140460968, + 0.021184707060456276, + 0.029575904831290245, + 0.012063912115991116, + 0.004895610269159079, + 0.009441059082746506, + 0.009991209954023361, + 0.005110329017043114, + 0.006104771979153156, + 0.00582580640912056, + 0.00636667013168335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -1.6006124019622803 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07908120006322861 + ], + "opposite_logprob_tokens": [ + -2.5790812969207764 + ], + "logprob_reward": -0.07908120006322861, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07908120006322861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20598258078098297, + 0.0866774246096611, + 0.029230020940303802, + 0.050823476165533066, + 0.047277841717004776, + 0.04565497860312462, + 0.047319114208221436, + 0.21310143172740936, + 0.20294471085071564, + 0.07098846882581711 + ], + "credit_map": [ + 0.023843659088015556, + 0.018279515206813812, + 0.0315815769135952, + 0.02227952890098095, + 0.03022533282637596, + 0.025602107867598534, + 0.02946329116821289, + 0.02470756322145462, + 0.013314399868249893, + 0.0048292698338627815, + 0.014729460701346397, + 0.01002613827586174, + 0.007391589228063822, + 0.021827194839715958, + 0.006655294448137283, + 0.007904077880084515, + 0.0034572193399071693, + 0.0017077084630727768, + 0.005481201224029064, + 0.003622264601290226, + 0.003335702233016491, + 0.004729447420686483, + 0.003096040105447173, + 0.003800438717007637, + 0.00632892269641161, + 0.00411248579621315, + 0.008540735580027103, + 0.005702691152691841, + 0.006720390636473894, + 0.007750577758997679, + 0.0057639675214886665, + 0.005903704557567835, + 0.004402357619255781, + 0.005343384575098753, + 0.005004477221518755, + 0.007583439815789461, + 0.007194293197244406, + 0.004807468969374895, + 0.005283787846565247, + 0.007658631540834904, + 0.004976837895810604, + 0.004531617276370525, + 0.005860285833477974, + 0.005645697936415672, + 0.006568806245923042, + 0.005462821573019028, + 0.006270759738981724, + 0.006338153965771198, + 0.00815051794052124, + 0.004932856187224388, + 0.004007978364825249, + 0.0046896375715732574, + 0.004404331557452679, + 0.010182032361626625, + 0.004130967427045107, + 0.006820794194936752, + 0.06652328372001648, + 0.011380827985703945, + 0.015830669552087784, + 0.013376814313232899, + 0.01409622561186552, + 0.0676463171839714, + 0.010819929651916027, + 0.013427354395389557, + 0.04335769638419151, + 0.01961543783545494, + 0.03127758204936981, + 0.03715138882398605, + 0.01604251191020012, + 0.02645670436322689, + 0.011854886077344418, + 0.017188498750329018, + 0.015176315791904926, + 0.006342930253595114, + 0.01045179646462202, + 0.011710644699633121, + 0.006303455680608749, + 0.006787121761590242, + 0.007068201899528503, + 0.0071480050683021545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.07908120006322861 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05501127615571022 + ], + "opposite_logprob_tokens": [ + -2.93001127243042 + ], + "logprob_reward": -0.05501127615571022, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05501127615571022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1965213119983673, + 0.0639304593205452, + 0.03879551962018013, + 0.05356796458363533, + 0.11684323102235794, + 0.17245234549045563, + 0.03990120440721512, + 0.10508590191602707, + 0.1430639922618866, + 0.06983799487352371 + ], + "credit_map": [ + 0.022311681881546974, + 0.018543051555752754, + 0.027203334495425224, + 0.024326365441083908, + 0.03032827563583851, + 0.023953886702656746, + 0.027909763157367706, + 0.02194494940340519, + 0.006626663263887167, + 0.005071169696748257, + 0.012016168795526028, + 0.010737385600805283, + 0.005935871973633766, + 0.00938754715025425, + 0.00674395402893424, + 0.0074116941541433334, + 0.004487669561058283, + 0.003697853535413742, + 0.005022724624723196, + 0.006898111663758755, + 0.003849975299090147, + 0.005476693622767925, + 0.0047815823927521706, + 0.004580909851938486, + 0.006884438451379538, + 0.004663500469177961, + 0.005573182832449675, + 0.007801794912666082, + 0.006056681275367737, + 0.010753915645182133, + 0.0050533609464764595, + 0.0067810905165970325, + 0.010214699432253838, + 0.012675780802965164, + 0.012447546236217022, + 0.016396423801779747, + 0.01795317605137825, + 0.013392305001616478, + 0.01570911891758442, + 0.018054183572530746, + 0.01809227652847767, + 0.01659032516181469, + 0.015629062429070473, + 0.019612912088632584, + 0.028134120628237724, + 0.02023959904909134, + 0.026513537392020226, + 0.027640506625175476, + 0.006500419694930315, + 0.003968340810388327, + 0.003907247446477413, + 0.004357172176241875, + 0.004376232158392668, + 0.00691570108756423, + 0.003800493897870183, + 0.0060755969025194645, + 0.021915676072239876, + 0.009979456663131714, + 0.012900125235319138, + 0.014934959821403027, + 0.008966261520981789, + 0.019915541633963585, + 0.008159909397363663, + 0.008313970640301704, + 0.019757039844989777, + 0.016513532027602196, + 0.021791916340589523, + 0.023129615932703018, + 0.013014346361160278, + 0.0154016287997365, + 0.014128300361335278, + 0.01932762935757637, + 0.015542884357273579, + 0.006247946061193943, + 0.008137964643537998, + 0.012285985983908176, + 0.006210022140294313, + 0.006506763398647308, + 0.007298355456441641, + 0.007608074229210615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.05501127615571022 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234209299087524 + ], + "opposite_logprob_tokens": [ + -0.8984209299087524 + ], + "logprob_reward": -0.5234209299087524, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234209299087524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.114692822098732, + 0.04277851805090904, + 0.039939675480127335, + 0.05848241597414017, + 0.058464065194129944, + 0.11429223418235779, + 0.09916824847459793, + 0.3083135783672333, + 0.1237126812338829, + 0.04015573486685753 + ], + "credit_map": [ + 0.011705576442182064, + 0.010565336793661118, + 0.021028950810432434, + 0.0164027102291584, + 0.012747107073664665, + 0.013816741295158863, + 0.012940368615090847, + 0.015486031770706177, + 0.005027156323194504, + 0.003629833459854126, + 0.007333796005696058, + 0.004644989501684904, + 0.005230866838246584, + 0.006483467295765877, + 0.005834418348968029, + 0.004593989811837673, + 0.004396605305373669, + 0.0042125023901462555, + 0.003241290571168065, + 0.0055937254801392555, + 0.004815889056771994, + 0.0068572815507650375, + 0.005820136982947588, + 0.005002241116017103, + 0.008669593371450901, + 0.005324681755155325, + 0.004819392692297697, + 0.007675220724195242, + 0.008647187612950802, + 0.010990995913743973, + 0.00647989334538579, + 0.005875447299331427, + 0.0056690131314098835, + 0.007347764912992716, + 0.006796092260628939, + 0.009279066696763039, + 0.008286253549158573, + 0.006949474569410086, + 0.006139247212558985, + 0.007997150532901287, + 0.011154967360198498, + 0.014071322977542877, + 0.01226950716227293, + 0.01653403788805008, + 0.017764568328857422, + 0.013218194246292114, + 0.015400439500808716, + 0.013879189267754555, + 0.013318019919097424, + 0.008082875981926918, + 0.016645673662424088, + 0.011109554208815098, + 0.013078504242002964, + 0.0131042106077075, + 0.013732323423027992, + 0.010097086429595947, + 0.10673686116933823, + 0.01394748967140913, + 0.018025485798716545, + 0.013127154670655727, + 0.017369244247674942, + 0.10694479942321777, + 0.011279549449682236, + 0.020882999524474144, + 0.01682738959789276, + 0.013416439294815063, + 0.015582159161567688, + 0.02175229787826538, + 0.012684240937232971, + 0.012477701529860497, + 0.018386557698249817, + 0.012585886754095554, + 0.005335990339517593, + 0.0033066372852772474, + 0.008061503060162067, + 0.007654669228941202, + 0.0037671199534088373, + 0.004683477804064751, + 0.0036982127930969, + 0.003648124635219574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.5234209299087524 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up, the camera remains still, its focus locked precisely on a new fern leaf as it slowly unfurls from its signature curl. A vibrant emerald green, covered in a dense layer of fine white hairs, the leaf appears vibrant in the soft, mist-diffused golden light of dawn. Precariously suspended at the tip of the leaf, a perfectly spherical dewdrop shimmers, its crystal-clear form like a miniature crystal, its limpid interior reflecting and distorting the out-of-focus forest backdrop, shimmering with light. As a barely perceptible breeze sweeps across the leaf, a subtle yet constant tremor ripples through the leaf, a movement so slight that it sways the dewdrop itself, causing light to flow steadily and deliberately across its surface. A thick, translucent morning mist gently swirls in the background, obscuring everything beyond, leaving only the soft hues of moss and damp soil, creating an atmosphere of profound tranquility and new hope.", + "seed": 26019, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_9334_seed_26019_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.19788322933018207, + "vlm_logprob": -0.19788322933018207, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.19788322933018207, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03591001033782959 + ], + "opposite_logprob_tokens": [ + -4.410910129547119 + ], + "logprob_reward": -0.03591001033782959, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03591001033782959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3410225510597229, + 0.11040697246789932, + 0.0866713598370552, + 0.09371481835842133, + 0.0361788272857666, + 0.04653190076351166, + 0.06355723738670349, + 0.1583860218524933, + 0.01552699413150549, + 0.04800327494740486 + ], + "credit_map": [ + 0.05828538164496422, + 0.009570767171680927, + 0.13704907894134521, + 0.02921115793287754, + 0.00823915284126997, + 0.01673365756869316, + 0.016522999852895737, + 0.06541034579277039, + 0.027298588305711746, + 0.014015184715390205, + 0.016996217891573906, + 0.012954425066709518, + 0.008418078534305096, + 0.007102595642209053, + 0.012489217333495617, + 0.011132671497762203, + 0.01966698095202446, + 0.004687171429395676, + 0.010478293523192406, + 0.01026964746415615, + 0.005644503515213728, + 0.0046472568064928055, + 0.01669338159263134, + 0.014584124088287354, + 0.026072075590491295, + 0.006454015616327524, + 0.010839360766112804, + 0.008419398218393326, + 0.004133984446525574, + 0.007003335282206535, + 0.009394820779561996, + 0.021397825330495834, + 0.009459998458623886, + 0.0026543177664279938, + 0.005737552884966135, + 0.003370301565155387, + 0.0021791744511574507, + 0.002947838045656681, + 0.0036146105267107487, + 0.006215034984052181, + 0.008587087504565716, + 0.004184178542345762, + 0.01029318105429411, + 0.0066980281844735146, + 0.0027535720728337765, + 0.002678387565538287, + 0.004951285198330879, + 0.00638618366792798, + 0.009298714809119701, + 0.0037043518386781216, + 0.02154899761080742, + 0.00931944977492094, + 0.0025805016048252583, + 0.0033024572767317295, + 0.004769711755216122, + 0.009033049456775188, + 0.05368728190660477, + 0.009391702711582184, + 0.04519077390432358, + 0.016550717875361443, + 0.004280474968254566, + 0.006395070347934961, + 0.0092534851282835, + 0.013636498712003231, + 0.004659922327846289, + 0.0007416597800329328, + 0.0023524418938905, + 0.0017986687598749995, + 0.0006684465915895998, + 0.0009245710680261254, + 0.0021825144067406654, + 0.0021987694781273603, + 0.010588792152702808, + 0.0023978550452739, + 0.008325918577611446, + 0.006712255068123341, + 0.0025882842019200325, + 0.0033076400868594646, + 0.0063864667899906635, + 0.007696060463786125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.03591001033782959 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298324167728424 + ], + "opposite_logprob_tokens": [ + -2.237983226776123 + ], + "logprob_reward": -0.11298324167728424, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298324167728424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47288331389427185, + 0.06177316606044769, + 0.09265553206205368, + 0.033754196017980576, + 0.03862128034234047, + 0.026102501899003983, + 0.03806295245885849, + 0.07359926402568817, + 0.10204237699508667, + 0.06050536036491394 + ], + "credit_map": [ + 0.07812618464231491, + 0.013519170694053173, + 0.19202610850334167, + 0.03693122789263725, + 0.010781243443489075, + 0.022777356207370758, + 0.026193775236606598, + 0.09252829104661942, + 0.014673174358904362, + 0.009359234943985939, + 0.006926740054041147, + 0.009609943255782127, + 0.0053871069103479385, + 0.004073352552950382, + 0.004948862362653017, + 0.0067947618663311005, + 0.01282037515193224, + 0.009594433940947056, + 0.01585080847144127, + 0.012827512808144093, + 0.005122336558997631, + 0.009912052191793919, + 0.0116965863853693, + 0.014831436797976494, + 0.006087171379476786, + 0.003374540712684393, + 0.0038348466623574495, + 0.004138812888413668, + 0.003058479633182287, + 0.003073000581935048, + 0.004633303266018629, + 0.005554045084863901, + 0.005266915075480938, + 0.004400060512125492, + 0.0068375589326024055, + 0.006918993778526783, + 0.003152577206492424, + 0.0035435056779533625, + 0.0037745044101029634, + 0.004727167077362537, + 0.005198812577873468, + 0.0022084906231611967, + 0.005650597158819437, + 0.0031090450938791037, + 0.0019115873146802187, + 0.0019715535454452038, + 0.002764751436188817, + 0.0032876641489565372, + 0.007709890138357878, + 0.002586273942142725, + 0.009699986316263676, + 0.005465335212647915, + 0.0015193376457318664, + 0.002109016990289092, + 0.003886840306222439, + 0.005086276680231094, + 0.019903292879462242, + 0.006079796236008406, + 0.012992178089916706, + 0.01032966934144497, + 0.0034650559537112713, + 0.006288065575063229, + 0.005665120668709278, + 0.008876094594597816, + 0.01583469659090042, + 0.007433650083839893, + 0.018173620104789734, + 0.015265466645359993, + 0.008220350369811058, + 0.00784805603325367, + 0.012581847608089447, + 0.016684703528881073, + 0.008729036897420883, + 0.005395970307290554, + 0.011365523561835289, + 0.007493575103580952, + 0.005649828817695379, + 0.0051865046843886375, + 0.007318270858377218, + 0.009366655722260475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.11298324167728424 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603561773896217 + ], + "opposite_logprob_tokens": [ + -3.895603656768799 + ], + "logprob_reward": -0.020603561773896217, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603561773896217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39992690086364746, + 0.10089755803346634, + 0.2701367139816284, + 0.024497410282492638, + 0.024826861917972565, + 0.0429924912750721, + 0.02996809035539627, + 0.02386583760380745, + 0.01911062002182007, + 0.06377750635147095 + ], + "credit_map": [ + 0.07004311680793762, + 0.013938432559370995, + 0.1624339520931244, + 0.051986224949359894, + 0.00865912064909935, + 0.017344916239380836, + 0.017920812591910362, + 0.05760028958320618, + 0.020614195615053177, + 0.010105274617671967, + 0.01688731648027897, + 0.015760742127895355, + 0.007856689393520355, + 0.006457106675952673, + 0.009455441497266293, + 0.013760779984295368, + 0.04994130879640579, + 0.022964177653193474, + 0.03459947556257248, + 0.033752888441085815, + 0.013887317851185799, + 0.019105849787592888, + 0.05100925266742706, + 0.04487641900777817, + 0.003501426661387086, + 0.0020284170750528574, + 0.004476140718907118, + 0.0031711093615740538, + 0.002087641740217805, + 0.0016349771758541465, + 0.0035585525911301374, + 0.004039143677800894, + 0.003965876996517181, + 0.00217005074955523, + 0.004587076138705015, + 0.003564714454114437, + 0.0023009239230304956, + 0.002916378667578101, + 0.002824165392667055, + 0.0024976744316518307, + 0.012927177362143993, + 0.0031951756682246923, + 0.007210193667560816, + 0.005167434457689524, + 0.002816972555592656, + 0.0027725971303880215, + 0.004426346626132727, + 0.004476591479033232, + 0.004153724759817123, + 0.0018642223440110683, + 0.009618081152439117, + 0.006722230929881334, + 0.0014862940879538655, + 0.0014512595953419805, + 0.0017919441452249885, + 0.0028803327586501837, + 0.005745534785091877, + 0.0015331238973885775, + 0.0056417775340378284, + 0.0031873569823801517, + 0.001118368934839964, + 0.0009959284216165543, + 0.0028644222766160965, + 0.002779322909191251, + 0.0034000680316239595, + 0.0016113172750920057, + 0.0031700076069682837, + 0.0028858205769211054, + 0.001477832905948162, + 0.0014786985702812672, + 0.00257993140257895, + 0.0025069417897611856, + 0.012285218574106693, + 0.0040855491533875465, + 0.011755397543311119, + 0.007532537914812565, + 0.003799079218879342, + 0.005751538556069136, + 0.007991916500031948, + 0.01057626772671938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.020603561773896217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062073156237602234 + ], + "opposite_logprob_tokens": [ + -2.812073230743408 + ], + "logprob_reward": -0.062073156237602234, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062073156237602234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10314583033323288, + 0.14583627879619598, + 0.5301903486251831, + 0.015134042128920555, + 0.033164843916893005, + 0.025921838358044624, + 0.031070882454514503, + 0.0232705045491457, + 0.04979611933231354, + 0.04246930405497551 + ], + "credit_map": [ + 0.014645958319306374, + 0.004270432982593775, + 0.03595151752233505, + 0.01016467995941639, + 0.004194736015051603, + 0.006238107569515705, + 0.006579733919352293, + 0.021100670099258423, + 0.038189467042684555, + 0.015560861676931381, + 0.017742911353707314, + 0.020186716690659523, + 0.011440473608672619, + 0.011065592058002949, + 0.012706891633570194, + 0.018943367525935173, + 0.11175300180912018, + 0.041336413472890854, + 0.07610727101564407, + 0.0686616450548172, + 0.024410242214798927, + 0.03839749097824097, + 0.08301039040088654, + 0.08651392161846161, + 0.0016979823121801019, + 0.0013157392386347055, + 0.002270722994580865, + 0.0031937798485159874, + 0.0012452395167201757, + 0.00114386563654989, + 0.0022638244554400444, + 0.0020028878934681416, + 0.006329088006168604, + 0.0037215149495750666, + 0.005435938015580177, + 0.005730871576815844, + 0.00212061102502048, + 0.003021363401785493, + 0.003142483765259385, + 0.0036629741080105305, + 0.0040140096098184586, + 0.0027768665459007025, + 0.0037378959823399782, + 0.004840980749577284, + 0.0019259696127846837, + 0.0025548201519995928, + 0.0034035667777061462, + 0.0026677276473492384, + 0.0056951758451759815, + 0.001896837493404746, + 0.010185539722442627, + 0.0029661289881914854, + 0.0017773561412468553, + 0.0015863421140238643, + 0.0028891197871416807, + 0.004074380733072758, + 0.004093159455806017, + 0.0016711162170395255, + 0.003796863369643688, + 0.00285018770955503, + 0.0014710047980770469, + 0.0019615611527115107, + 0.0030201224144548178, + 0.00440648989751935, + 0.008149501867592335, + 0.0042320662178099155, + 0.008149656467139721, + 0.006677751895040274, + 0.004764629993587732, + 0.004447168204933405, + 0.005597312934696674, + 0.007778033148497343, + 0.009423450566828251, + 0.0024318716023117304, + 0.0063577573746442795, + 0.005338695365935564, + 0.002169993007555604, + 0.0041382028721272945, + 0.005106346681714058, + 0.007502988446503878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.062073156237602234 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578461766242981 + ], + "opposite_logprob_tokens": [ + -0.6328461766242981 + ], + "logprob_reward": -0.7578461766242981, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578461766242981, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.25260964035987854, + 0.07804150134325027, + 0.17836891114711761, + 0.05322658270597458, + 0.047348469495773315, + 0.09060803800821304, + 0.04082626476883888, + 0.10151377320289612, + 0.1384471356868744, + 0.019009707495570183 + ], + "credit_map": [ + 0.029619082808494568, + 0.0076119182631373405, + 0.08695302903652191, + 0.0242350734770298, + 0.005684263072907925, + 0.01735711842775345, + 0.014694096520543098, + 0.06645506620407104, + 0.021286126226186752, + 0.010265265591442585, + 0.010249226354062557, + 0.008299405686557293, + 0.005239853635430336, + 0.0049131885170936584, + 0.008016855455935001, + 0.009771579876542091, + 0.03328205272555351, + 0.013623230159282684, + 0.020819561555981636, + 0.028176244348287582, + 0.009196214377880096, + 0.010399837046861649, + 0.034605249762535095, + 0.02826652303338051, + 0.010568702593445778, + 0.004409785382449627, + 0.0077185677364468575, + 0.005350357387214899, + 0.0042358944192528725, + 0.00412818742915988, + 0.005973156075924635, + 0.010841929353773594, + 0.009996836073696613, + 0.005586269311606884, + 0.007113284431397915, + 0.007509282790124416, + 0.002341725630685687, + 0.004249612335115671, + 0.0050285677425563335, + 0.005522888153791428, + 0.018658967688679695, + 0.009774791076779366, + 0.013110579922795296, + 0.013841127045452595, + 0.0034966417588293552, + 0.008596827276051044, + 0.007754203397780657, + 0.015374901704490185, + 0.01263455394655466, + 0.002166480990126729, + 0.013706136494874954, + 0.0034050254616886377, + 0.0012783755082637072, + 0.0016469231341034174, + 0.002634454518556595, + 0.003354314249008894, + 0.013207140378654003, + 0.0061992560513317585, + 0.031248142942786217, + 0.008237969130277634, + 0.005648783873766661, + 0.006412984803318977, + 0.008182931691408157, + 0.022376572713255882, + 0.02256147377192974, + 0.012104549445211887, + 0.022189026698470116, + 0.018999062478542328, + 0.013750804588198662, + 0.01230579148977995, + 0.015580175444483757, + 0.0209562536329031, + 0.0040373243391513824, + 0.0013963133096694946, + 0.003417109837755561, + 0.002069678856059909, + 0.001195531920529902, + 0.001614337437786162, + 0.0022678526584059, + 0.003011559136211872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.7578461766242981 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "In an extreme macro close-up, the camera remains still, its focus locked precisely on a new fern leaf as it slowly unfurls from its signature curl. A vibrant emerald green, covered in a dense layer of fine white hairs, the leaf appears vibrant in the soft, mist-diffused golden light of dawn. Precariously suspended at the tip of the leaf, a perfectly spherical dewdrop shimmers, its crystal-clear form like a miniature crystal, its limpid interior reflecting and distorting the out-of-focus forest backdrop, shimmering with light. As a barely perceptible breeze sweeps across the leaf, a subtle yet constant tremor ripples through the leaf, a movement so slight that it sways the dewdrop itself, causing light to flow steadily and deliberately across its surface. A thick, translucent morning mist gently swirls in the background, obscuring everything beyond, leaving only the soft hues of moss and damp soil, creating an atmosphere of profound tranquility and new hope.", + "seed": 26020, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_7/index_9334_seed_26020_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09130658705253154, + "vlm_logprob": -0.09130658705253154, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.09130658705253154, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03041601926088333 + ], + "opposite_logprob_tokens": [ + -4.655416011810303 + ], + "logprob_reward": -0.03041601926088333, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.03041601926088333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04152805730700493, + 0.08015257120132446, + 0.018725043162703514, + 0.056806933134794235, + 0.26013725996017456, + 0.06669074296951294, + 0.3193785846233368, + 0.04578590393066406, + 0.01939481869339943, + 0.09140006452798843 + ], + "credit_map": [ + 0.004117444157600403, + 0.0036330013535916805, + 0.006019409280270338, + 0.006473954301327467, + 0.004297131672501564, + 0.006009627599269152, + 0.004482005722820759, + 0.006495488341897726, + 0.01208474487066269, + 0.004486722405999899, + 0.02586495503783226, + 0.0072427792474627495, + 0.004967868328094482, + 0.00553172966465354, + 0.005792086012661457, + 0.01418168656527996, + 0.0026867594569921494, + 0.00220174971036613, + 0.002305007306858897, + 0.002465531462803483, + 0.002082176273688674, + 0.0017741628689691424, + 0.0030626736115664244, + 0.002146983752027154, + 0.012051304802298546, + 0.0034496879670768976, + 0.020752856507897377, + 0.0033983541652560234, + 0.0019473839784041047, + 0.0025396174751222134, + 0.0020981424022465944, + 0.010569588281214237, + 0.06224902346730232, + 0.015111393295228481, + 0.09911169111728668, + 0.01499014813452959, + 0.006832754239439964, + 0.009204479865729809, + 0.006853987462818623, + 0.04578377678990364, + 0.01554088480770588, + 0.004333363380283117, + 0.022273417562246323, + 0.004840917885303497, + 0.0031097230967134237, + 0.0027041444554924965, + 0.004385609179735184, + 0.009502686560153961, + 0.080597884953022, + 0.018748879432678223, + 0.1207563579082489, + 0.013136787340044975, + 0.00982364360243082, + 0.009696620516479015, + 0.008908899500966072, + 0.05770953372120857, + 0.01028669998049736, + 0.004117179196327925, + 0.010997897014021873, + 0.0034877716097980738, + 0.0035472961608320475, + 0.0036167558282613754, + 0.004007334820926189, + 0.005724974442273378, + 0.0016643768176436424, + 0.0012747258879244328, + 0.005381432361900806, + 0.0015339631354436278, + 0.0016564078396186233, + 0.0018152936827391386, + 0.0015303242253139615, + 0.004538295790553093, + 0.011314036324620247, + 0.0040812077932059765, + 0.03187841176986694, + 0.007616821210831404, + 0.004551276098936796, + 0.005755370948463678, + 0.0057555390521883965, + 0.020447412505745888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.03041601926088333 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428365260362625 + ], + "opposite_logprob_tokens": [ + -2.017836570739746 + ], + "logprob_reward": -0.1428365260362625, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428365260362625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061683427542448044, + 0.14710767567157745, + 0.05536942183971405, + 0.04711364954710007, + 0.07149898260831833, + 0.031687140464782715, + 0.3764389157295227, + 0.06384164839982986, + 0.11558276414871216, + 0.029676388949155807 + ], + "credit_map": [ + 0.015068834647536278, + 0.008624441921710968, + 0.008770045824348927, + 0.008048094809055328, + 0.004636913537979126, + 0.005327222868800163, + 0.00565794762223959, + 0.005549929104745388, + 0.028873644769191742, + 0.009454059414565563, + 0.05616190657019615, + 0.007267334032803774, + 0.0038346555083990097, + 0.0054886434227228165, + 0.006387065630406141, + 0.029640359804034233, + 0.013204392045736313, + 0.006322075612843037, + 0.012948451563715935, + 0.00393063435330987, + 0.0033595452550798655, + 0.0037726801820099354, + 0.0050538708455860615, + 0.00677777174860239, + 0.00787401758134365, + 0.0025919650215655565, + 0.017192630097270012, + 0.002440188778564334, + 0.001887159189209342, + 0.0023469850420951843, + 0.0024835187941789627, + 0.010297185741364956, + 0.008665064349770546, + 0.004006583709269762, + 0.019718803465366364, + 0.005878761876374483, + 0.00519930012524128, + 0.003914493136107922, + 0.012390988878905773, + 0.011724988929927349, + 0.0025674495846033096, + 0.00337964971549809, + 0.0044909510761499405, + 0.005762105342000723, + 0.0034958007745444775, + 0.0038455079775303602, + 0.0039863972924649715, + 0.00415927916765213, + 0.09623197466135025, + 0.022689364850521088, + 0.14302973449230194, + 0.01592722348868847, + 0.011345402337610722, + 0.010108201764523983, + 0.00996888242661953, + 0.06713816523551941, + 0.00892847403883934, + 0.006188275758177042, + 0.009298804216086864, + 0.008239926770329475, + 0.0062712146900594234, + 0.005590591114014387, + 0.00782057549804449, + 0.011503783985972404, + 0.012961072847247124, + 0.004853021819144487, + 0.039904799312353134, + 0.007395675405859947, + 0.004195860121399164, + 0.008371841162443161, + 0.007519462611526251, + 0.03038102574646473, + 0.003265128005295992, + 0.0014934458304196596, + 0.008847799152135849, + 0.002739547286182642, + 0.0020590946078300476, + 0.002382489386945963, + 0.0026934724301099777, + 0.006195411551743746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.1428365260362625 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816141352057457 + ], + "opposite_logprob_tokens": [ + -3.529816150665283 + ], + "logprob_reward": -0.029816141352057457, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816141352057457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09095205366611481, + 0.22136607766151428, + 0.009814513847231865, + 0.018430616706609726, + 0.043177731335163116, + 0.013772605918347836, + 0.49917981028556824, + 0.07497850060462952, + 0.020993690937757492, + 0.007334388792514801 + ], + "credit_map": [ + 0.018389908596873283, + 0.006901715416461229, + 0.032153382897377014, + 0.008021955378353596, + 0.0030234456062316895, + 0.004212516359984875, + 0.003928649704903364, + 0.01432048063725233, + 0.050939083099365234, + 0.01880437321960926, + 0.08078260719776154, + 0.010987767018377781, + 0.006536767352372408, + 0.00750112347304821, + 0.006764194928109646, + 0.039050180464982986, + 0.0008038152591325343, + 0.0009762512636370957, + 0.0012182930950075388, + 0.0013283230364322662, + 0.0014572353102266788, + 0.0011065207654610276, + 0.0016161738894879818, + 0.0013079020427539945, + 0.001988275209441781, + 0.0012129070237278938, + 0.004991841036826372, + 0.0017650909721851349, + 0.0013968298444524407, + 0.0013907450484111905, + 0.002012886106967926, + 0.003672041930258274, + 0.002954163122922182, + 0.003027872182428837, + 0.007706737145781517, + 0.005935792811214924, + 0.003806569380685687, + 0.0035891272127628326, + 0.009396739304065704, + 0.006760736927390099, + 0.002416044706478715, + 0.0008261895272880793, + 0.0039885807782411575, + 0.0015371269546449184, + 0.000988248037174344, + 0.0008514758083038032, + 0.0009362517739646137, + 0.0022286903113126755, + 0.12641900777816772, + 0.02919299714267254, + 0.19138117134571075, + 0.020423199981451035, + 0.01467309333384037, + 0.01334979385137558, + 0.013017275370657444, + 0.09072329103946686, + 0.0174461230635643, + 0.004444032441824675, + 0.0251923855394125, + 0.0036471730563789606, + 0.0038767163641750813, + 0.0031401130836457014, + 0.004242731723934412, + 0.01298922672867775, + 0.0016400025924667716, + 0.0010873524006456137, + 0.00633294926956296, + 0.001488362904638052, + 0.0008563409210182726, + 0.0018611811101436615, + 0.0019762085285037756, + 0.005751292686909437, + 0.0009185241651721299, + 0.000581375032197684, + 0.001944037969224155, + 0.0006336956284940243, + 0.0006680244696326554, + 0.0007730295765213668, + 0.0005362102529034019, + 0.0012794925132766366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.029816141352057457 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013392536202445626 + ], + "opposite_logprob_tokens": [ + -6.626339435577393 + ], + "logprob_reward": -0.0013392536202445626, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013392536202445626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2153656929731369, + 0.1221025213599205, + 0.014242159202694893, + 0.0382491797208786, + 0.10355244576931, + 0.023640146479010582, + 0.32002493739128113, + 0.099964939057827, + 0.04754103720188141, + 0.015316958539187908 + ], + "credit_map": [ + 0.05062224343419075, + 0.01260224636644125, + 0.08584818989038467, + 0.011832775548100471, + 0.005359449423849583, + 0.007870281115174294, + 0.0072142076678574085, + 0.034016285091638565, + 0.030285706743597984, + 0.011185944080352783, + 0.04448240250349045, + 0.005938443820923567, + 0.002820369554683566, + 0.003398441942408681, + 0.003279914613813162, + 0.020711280405521393, + 0.0031004645861685276, + 0.0011073561618104577, + 0.003314429894089699, + 0.0011776791652664542, + 0.0013029633555561304, + 0.001037648762576282, + 0.001585452351719141, + 0.0016161642270162702, + 0.004041280597448349, + 0.0021514163818210363, + 0.011040261015295982, + 0.003270047949627042, + 0.0027903467416763306, + 0.0029166853055357933, + 0.00423714192584157, + 0.007801995612680912, + 0.022294359281659126, + 0.005516823381185532, + 0.03535449504852295, + 0.008385170251131058, + 0.004445620346814394, + 0.0052450550720095634, + 0.005506530869752169, + 0.01680438593029976, + 0.0023172318469733, + 0.0026770757976919413, + 0.002652803435921669, + 0.0047407811507582664, + 0.0026802578940987587, + 0.0023693260736763477, + 0.003462099237367511, + 0.0027405680157244205, + 0.08046410232782364, + 0.018597077578306198, + 0.12266279011964798, + 0.01296421978622675, + 0.00934791099280119, + 0.008932150900363922, + 0.008566074073314667, + 0.0584905743598938, + 0.018064018338918686, + 0.004747744649648666, + 0.035273227840662, + 0.0055829007178545, + 0.004285918083041906, + 0.005017980467528105, + 0.005606187507510185, + 0.02138696238398552, + 0.005217427387833595, + 0.002269771648570895, + 0.01606230065226555, + 0.0031583684030920267, + 0.0021517074201256037, + 0.0032291412353515625, + 0.00380569975823164, + 0.011646616272628307, + 0.0016225081635639071, + 0.0008982468862086535, + 0.004967750050127506, + 0.001287255436182022, + 0.0009899279102683067, + 0.0007907094550319016, + 0.0014083527494221926, + 0.003352206898853183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0013392536202445626 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up with the camera remaining still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a new fern leaf that is slowly unfurling from a curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leaf vibrant emerald green and covered in a dense layer of fine white hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a perfectly spherical, crystal-clear dewdrop suspended at the tip of the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the leaf exhibit a subtle tremor caused by a breeze that sways the dewdrop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25212499499320984 + ], + "opposite_logprob_tokens": [ + -1.5021250247955322 + ], + "logprob_reward": -0.25212499499320984, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25212499499320984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15119968354701996, + 0.3154924511909485, + 0.0397409088909626, + 0.014455810189247131, + 0.19501037895679474, + 0.018155720084905624, + 0.1126372292637825, + 0.08507802337408066, + 0.051023684442043304, + 0.017206139862537384 + ], + "credit_map": [ + 0.035814885050058365, + 0.008764687925577164, + 0.05916990339756012, + 0.009177207946777344, + 0.0041794064454734325, + 0.006200357805937529, + 0.005680684931576252, + 0.02221253141760826, + 0.06676570326089859, + 0.02345963940024376, + 0.11686249822378159, + 0.016849137842655182, + 0.010052717290818691, + 0.01146015152335167, + 0.01107888761907816, + 0.05896368250250816, + 0.006430171430110931, + 0.0021182082127779722, + 0.013957157731056213, + 0.003237395780161023, + 0.002146091079339385, + 0.002113818656653166, + 0.002726120874285698, + 0.00701194116845727, + 0.003788574133068323, + 0.0012647821567952633, + 0.00364891579374671, + 0.00122169335372746, + 0.0009850425412878394, + 0.0009625076781958342, + 0.0009359765681438148, + 0.0016483162762597203, + 0.04789208620786667, + 0.012070340104401112, + 0.07370658963918686, + 0.010384639725089073, + 0.005094538908451796, + 0.006426695734262466, + 0.005977678578346968, + 0.033457789570093155, + 0.0018233859445899725, + 0.0016882129712030292, + 0.003317424561828375, + 0.003083593677729368, + 0.0016124940011650324, + 0.0016759374411776662, + 0.0024112265091389418, + 0.002543442649766803, + 0.028379041701555252, + 0.006519608665257692, + 0.04241623729467392, + 0.004515927750617266, + 0.0035249192733317614, + 0.003576024202629924, + 0.003418936161324382, + 0.020286530256271362, + 0.022284574806690216, + 0.007407812867313623, + 0.018777670338749886, + 0.005969993304461241, + 0.006046602968126535, + 0.006179219577461481, + 0.008817845024168491, + 0.009594300761818886, + 0.004878149833530188, + 0.0032026234548538923, + 0.013070388697087765, + 0.0032080819364637136, + 0.0029107537120580673, + 0.00456521101295948, + 0.005313451401889324, + 0.0138750234618783, + 0.001427158946171403, + 0.0015307387802749872, + 0.0028404828626662493, + 0.001556446892209351, + 0.002494239481166005, + 0.002854378195479512, + 0.001582728116773069, + 0.00291996356099844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.25212499499320984 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_8/index_tensor([4328, 4328])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_8/index_tensor([4328, 4328])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1686eb0ab92cc11fe457e9b350542e20cc694c47 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_8/index_tensor([4328, 4328])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The shot begins with an extreme close-up of a brilliantly colored hummingbird, its emerald green feathers gleaming metallically in the soft golden morning light. Its wings vibrate at a rate too rapid to detect with the naked eye as it hovers in mid-air. It carefully dips its needle-thin beak into the depths of a trumpet-shaped, scarlet flower, its petals still shimmering with morning dew. Then, the camera smoothly pulls back, gradually unfolding this microscopic canvas of life, revealing the flower growing in a vibrant green valley. A few quaint wooden cottages dot the valley floor, covered in lush grass and wildflowers, with plumes of smoke rising from their rooftops. The camera continues to steadily and calmly pull back, the field of view becomes increasingly wider, and finally freezes on a magnificent vista, showing the majestic and rugged mountains surrounding the entire idyllic valley. The outlines of the peaks are looming in the morning mist. The tones of the whole picture are warm and fresh, creating a tranquil yet shocking visual experience from the delicacy of microscopic life to the magnificence of macro nature.", + "seed": 4480, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_8/index_4328_seed_4480_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.010444643488153816, + "vlm_logprob": -0.010444643488153816, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.010444643488153816, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010841751471161842 + ], + "opposite_logprob_tokens": [ + -4.7608418464660645 + ], + "logprob_reward": -0.010841751471161842, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010841751471161842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4593878984451294, + 0.055525172501802444, + 0.0860799178481102, + 0.08860171586275101, + 0.08105023205280304, + 0.056815821677446365, + 0.08163747191429138, + 0.026764165610074997, + 0.014880958944559097, + 0.04925663769245148 + ], + "credit_map": [ + 0.08146732300519943, + 0.06812744587659836, + 0.09149656444787979, + 0.025330329313874245, + 0.03337559476494789, + 0.07325166463851929, + 0.032465722411870956, + 0.05387322977185249, + 0.006769245956093073, + 0.005006894003599882, + 0.005223049316555262, + 0.006135962903499603, + 0.008211789652705193, + 0.011075335554778576, + 0.005794139578938484, + 0.007308749482035637, + 0.010436799377202988, + 0.015487520024180412, + 0.014186877757310867, + 0.00846282672137022, + 0.010403038002550602, + 0.010347874835133553, + 0.006275498773902655, + 0.01047947071492672, + 0.006971149239689112, + 0.014625591225922108, + 0.008048239164054394, + 0.010969838127493858, + 0.009555766358971596, + 0.007807943504303694, + 0.018939362838864326, + 0.011683817952871323, + 0.010732670314610004, + 0.007332297042012215, + 0.013291001319885254, + 0.011402959004044533, + 0.00787531677633524, + 0.011325993575155735, + 0.007963874377310276, + 0.011126110330224037, + 0.007869587279856205, + 0.009564699605107307, + 0.007422362919896841, + 0.007155485451221466, + 0.004829855635762215, + 0.006499598268419504, + 0.006266150623559952, + 0.007208077237010002, + 0.006891826167702675, + 0.010011718608438969, + 0.011654231697320938, + 0.017301220446825027, + 0.0061980667524039745, + 0.011309594847261906, + 0.007786700036376715, + 0.01048410777002573, + 0.0019915751181542873, + 0.0019362192833796144, + 0.002830941928550601, + 0.003526404034346342, + 0.0025642483960837126, + 0.0028514540754258633, + 0.007155494764447212, + 0.003907825797796249, + 0.001096450723707676, + 0.0008882153197191656, + 0.0016348216449841857, + 0.0014805756509304047, + 0.001719857333227992, + 0.0018639768240973353, + 0.004082800820469856, + 0.00211425987072289, + 0.0033028624020516872, + 0.0038392217829823494, + 0.004495084285736084, + 0.0069068484008312225, + 0.004428790416568518, + 0.004036195110529661, + 0.0155715337023139, + 0.006676096934825182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.010841751471161842 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649423528462648 + ], + "opposite_logprob_tokens": [ + -5.379649639129639 + ], + "logprob_reward": -0.004649423528462648, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649423528462648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1965971738100052, + 0.07734828442335129, + 0.30381864309310913, + 0.09718173742294312, + 0.07656356692314148, + 0.07605841755867004, + 0.03437821567058563, + 0.03568968549370766, + 0.05510777235031128, + 0.047256533056497574 + ], + "credit_map": [ + 0.032807763665914536, + 0.025738729164004326, + 0.034580864012241364, + 0.011508471332490444, + 0.022573567926883698, + 0.03312326967716217, + 0.012940085493028164, + 0.023324405774474144, + 0.0064244261011481285, + 0.006831237580627203, + 0.009623710066080093, + 0.008949220180511475, + 0.016372213140130043, + 0.01082127820700407, + 0.008457868359982967, + 0.009868334047496319, + 0.03189868852496147, + 0.042171042412519455, + 0.036453064531087875, + 0.026991600170731544, + 0.02186954766511917, + 0.07328693568706512, + 0.027876392006874084, + 0.043271392583847046, + 0.012005248107016087, + 0.01586918905377388, + 0.006846111733466387, + 0.01193358562886715, + 0.011436461471021175, + 0.008191166445612907, + 0.018425649031996727, + 0.01247432641685009, + 0.00961330160498619, + 0.009468450210988522, + 0.011664420366287231, + 0.010357645340263844, + 0.00931510329246521, + 0.011867880821228027, + 0.005720064509660006, + 0.008556706830859184, + 0.01153834629803896, + 0.016553429886698723, + 0.010187572799623013, + 0.0077438196167349815, + 0.007134085986763239, + 0.007671578321605921, + 0.007204241584986448, + 0.00802534632384777, + 0.003999929875135422, + 0.0053610727190971375, + 0.004376342985779047, + 0.004824986681342125, + 0.0033651257399469614, + 0.004910672549158335, + 0.0033722470980137587, + 0.004167838022112846, + 0.0026664447505027056, + 0.0030965893529355526, + 0.0038932948373258114, + 0.0043104845099151134, + 0.00564924580976367, + 0.0032715103588998318, + 0.008098205551505089, + 0.00470391009002924, + 0.005232576746493578, + 0.005355457775294781, + 0.004909009672701359, + 0.005775555036962032, + 0.004790659062564373, + 0.00795045681297779, + 0.013553864322602749, + 0.00754019059240818, + 0.006727905478328466, + 0.007526089437305927, + 0.004528301768004894, + 0.004802653566002846, + 0.0056340983137488365, + 0.0055228774435818195, + 0.0076553416438400745, + 0.004859269130975008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.004649423528462648 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024931321386247873 + ], + "opposite_logprob_tokens": [ + -6.002492904663086 + ], + "logprob_reward": -0.0024931321386247873, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024931321386247873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08248688280582428, + 0.0775505006313324, + 0.09173743426799774, + 0.08859501034021378, + 0.1355229765176773, + 0.10893037170171738, + 0.19111409783363342, + 0.04467466101050377, + 0.05474504455924034, + 0.12464302033185959 + ], + "credit_map": [ + 0.014664717018604279, + 0.009173227474093437, + 0.01394317764788866, + 0.006380355451256037, + 0.00841391459107399, + 0.011279016733169556, + 0.009028066881000996, + 0.009604404680430889, + 0.006953034084290266, + 0.010066202841699123, + 0.009642056189477444, + 0.008542500436306, + 0.010496816597878933, + 0.01455453597009182, + 0.006432359106838703, + 0.010862992145121098, + 0.012614976614713669, + 0.016323743388056755, + 0.011810085736215115, + 0.012056496925652027, + 0.008011659607291222, + 0.011277908459305763, + 0.008977821096777916, + 0.010664744302630424, + 0.007058846298605204, + 0.015614541247487068, + 0.009915361180901527, + 0.010018030181527138, + 0.01360234897583723, + 0.009934142231941223, + 0.011636477895081043, + 0.010815264657139778, + 0.014636202715337276, + 0.010546647943556309, + 0.028863199055194855, + 0.025268081575632095, + 0.00866289995610714, + 0.019455531612038612, + 0.010417116805911064, + 0.017673300579190254, + 0.017885565757751465, + 0.024515625089406967, + 0.010780367068946362, + 0.013916173949837685, + 0.009532940573990345, + 0.011029612272977829, + 0.00975507777184248, + 0.01151501014828682, + 0.013767281547188759, + 0.017766902223229408, + 0.038873352110385895, + 0.04304511100053787, + 0.016258904710412025, + 0.01814216561615467, + 0.02039767988026142, + 0.022862685844302177, + 0.004555347375571728, + 0.0039426181465387344, + 0.005865989252924919, + 0.006357836537063122, + 0.0040930211544036865, + 0.005472590681165457, + 0.008933966979384422, + 0.005453288089483976, + 0.00628045666962862, + 0.0048660091124475, + 0.005145538132637739, + 0.007234353106468916, + 0.005365623161196709, + 0.005112228915095329, + 0.013420311734080315, + 0.00732052605599165, + 0.014934557490050793, + 0.01283701416105032, + 0.01077340729534626, + 0.015372908674180508, + 0.021735765039920807, + 0.014986199326813221, + 0.019508173689246178, + 0.014494990929961205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.0024931321386247873 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818063110113144 + ], + "opposite_logprob_tokens": [ + -4.018180847167969 + ], + "logprob_reward": -0.01818063110113144, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818063110113144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07902812957763672, + 0.04002576321363449, + 0.05142933502793312, + 0.09555276483297348, + 0.07696156203746796, + 0.039338368922472, + 0.218511164188385, + 0.049469150602817535, + 0.12503093481063843, + 0.22465281188488007 + ], + "credit_map": [ + 0.0163660179823637, + 0.008529488928616047, + 0.01087113842368126, + 0.008766872808337212, + 0.007713601924479008, + 0.010858256369829178, + 0.0068701691925525665, + 0.009052583947777748, + 0.004126355983316898, + 0.0050013246946036816, + 0.005128839518874884, + 0.005702816415578127, + 0.005163984838873148, + 0.005396696273237467, + 0.003945119678974152, + 0.00556063000112772, + 0.006402280647307634, + 0.007122046779841185, + 0.005454352591186762, + 0.0050102174282073975, + 0.005472646560519934, + 0.010624359361827374, + 0.004141738638281822, + 0.007201692089438438, + 0.007351084612309933, + 0.014980774372816086, + 0.00803268514573574, + 0.012470940127968788, + 0.011774340644478798, + 0.008308246731758118, + 0.020465277135372162, + 0.012169413268566132, + 0.015937892720103264, + 0.007703232113271952, + 0.011091318912804127, + 0.006945624947547913, + 0.00439932756125927, + 0.013640841469168663, + 0.0049585807137191296, + 0.012284742668271065, + 0.004568459931761026, + 0.006265247240662575, + 0.0036965198814868927, + 0.006397153250873089, + 0.0038378105964511633, + 0.005219958256930113, + 0.004360266495496035, + 0.0049929567612707615, + 0.0150376558303833, + 0.018768852576613426, + 0.04162042960524559, + 0.054700396955013275, + 0.016388531774282455, + 0.018953748047351837, + 0.024272430688142776, + 0.02876911871135235, + 0.004480788018554449, + 0.003721265820786357, + 0.00927506573498249, + 0.006904934998601675, + 0.0057965838350355625, + 0.0053424532525241375, + 0.0077301571145653725, + 0.00621790112927556, + 0.01515677385032177, + 0.012255067937076092, + 0.014940910041332245, + 0.019387558102607727, + 0.0111536281183362, + 0.017726704478263855, + 0.016749145463109016, + 0.017661146819591522, + 0.022660063579678535, + 0.023217950016260147, + 0.024396229535341263, + 0.03832915797829628, + 0.03244427219033241, + 0.019450681284070015, + 0.03667847439646721, + 0.027475964277982712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.01818063110113144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605827920138836 + ], + "opposite_logprob_tokens": [ + -4.141058444976807 + ], + "logprob_reward": -0.01605827920138836, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605827920138836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13556790351867676, + 0.11980722844600677, + 0.13798806071281433, + 0.17448779940605164, + 0.09518983960151672, + 0.09897755831480026, + 0.09220556914806366, + 0.06582162529230118, + 0.02647840790450573, + 0.05347592756152153 + ], + "credit_map": [ + 0.020605919882655144, + 0.015336166135966778, + 0.027777595445513725, + 0.0122641921043396, + 0.014739403501152992, + 0.019670333713293076, + 0.009747299365699291, + 0.015427010133862495, + 0.018468063324689865, + 0.017788685858249664, + 0.017152860760688782, + 0.012021365575492382, + 0.011693380773067474, + 0.015021118335425854, + 0.012704038061201572, + 0.014957727864384651, + 0.017908181995153427, + 0.016452860087156296, + 0.014370646327733994, + 0.017289770767092705, + 0.019547704607248306, + 0.01493646577000618, + 0.0221200343221426, + 0.015362413600087166, + 0.018133098259568214, + 0.027223508805036545, + 0.016237745061516762, + 0.017208121716976166, + 0.025734413415193558, + 0.02199181355535984, + 0.02643512189388275, + 0.021523991599678993, + 0.017581898719072342, + 0.007121096830815077, + 0.014728584326803684, + 0.016117002815008163, + 0.005773152224719524, + 0.012837383896112442, + 0.008657374419271946, + 0.01237335056066513, + 0.011674134992063046, + 0.01588151790201664, + 0.009534779004752636, + 0.015228372067213058, + 0.007914643734693527, + 0.01448782254010439, + 0.011899115517735481, + 0.012357190251350403, + 0.007849235087633133, + 0.012086779810488224, + 0.015774009749293327, + 0.02064766362309456, + 0.0067373765632510185, + 0.007971126586198807, + 0.010417274199426174, + 0.010722109116613865, + 0.006967486813664436, + 0.00968670379370451, + 0.010309997946023941, + 0.007818803191184998, + 0.007789190858602524, + 0.009069995023310184, + 0.007037795148789883, + 0.0071416618302464485, + 0.0026470855809748173, + 0.0027991647366434336, + 0.003484271466732025, + 0.004398546647280455, + 0.002677622716873884, + 0.0029291745740920305, + 0.00450099678710103, + 0.0030415495857596397, + 0.00506417453289032, + 0.00662181805819273, + 0.004199891816824675, + 0.007136677857488394, + 0.006783144548535347, + 0.007418472785502672, + 0.009852210991084576, + 0.006399544887244701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.01605827920138836 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with an extreme close-up of a brilliantly colored hummingbird, its emerald green feathers gleaming metallically in the soft golden morning light. Its wings vibrate at a rate too rapid to detect with the naked eye as it hovers in mid-air. It carefully dips its needle-thin beak into the depths of a trumpet-shaped, scarlet flower, its petals still shimmering with morning dew. Then, the camera smoothly pulls back, gradually unfolding this microscopic canvas of life, revealing the flower growing in a vibrant green valley. A few quaint wooden cottages dot the valley floor, covered in lush grass and wildflowers, with plumes of smoke rising from their rooftops. The camera continues to steadily and calmly pull back, the field of view becomes increasingly wider, and finally freezes on a magnificent vista, showing the majestic and rugged mountains surrounding the entire idyllic valley. The outlines of the peaks are looming in the morning mist. The tones of the whole picture are warm and fresh, creating a tranquil yet shocking visual experience from the delicacy of microscopic life to the magnificence of macro nature.", + "seed": 4481, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_8/index_4328_seed_4481_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.006170896673575043, + "vlm_logprob": -0.006170896673575043, + "vlm_yes_no_margin": 5.4, + "vlm_token_credit": -0.006170896673575043, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004692020360380411 + ], + "opposite_logprob_tokens": [ + -6.004692077636719 + ], + "logprob_reward": -0.004692020360380411, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.004692020360380411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1733090579509735, + 0.11124511063098907, + 0.07116540521383286, + 0.07393784821033478, + 0.32968685030937195, + 0.05490005761384964, + 0.07104477286338806, + 0.06759665906429291, + 0.018096407875418663, + 0.02901788055896759 + ], + "credit_map": [ + 0.022424954921007156, + 0.03694155439734459, + 0.018032092601060867, + 0.014180910773575306, + 0.025207843631505966, + 0.02493840828537941, + 0.01815897226333618, + 0.013424308970570564, + 0.016664253547787666, + 0.015690937638282776, + 0.01908719167113304, + 0.010446367785334587, + 0.007097289897501469, + 0.020201774314045906, + 0.009167131036520004, + 0.012890156358480453, + 0.006230627652257681, + 0.004446756094694138, + 0.023034904152154922, + 0.006470487453043461, + 0.004225084558129311, + 0.007711602374911308, + 0.004345444031059742, + 0.014700498431921005, + 0.011440531350672245, + 0.005082318093627691, + 0.010847734287381172, + 0.01046562846750021, + 0.00588852958753705, + 0.009575187228620052, + 0.009195898659527302, + 0.011442020535469055, + 0.08194516599178314, + 0.11653199791908264, + 0.024036111310124397, + 0.012992591597139835, + 0.015095897950232029, + 0.03133222833275795, + 0.014082759618759155, + 0.0336700938642025, + 0.005598783493041992, + 0.007270060013979673, + 0.006066482048481703, + 0.0042475746013224125, + 0.008552605286240578, + 0.008538477122783661, + 0.009898973628878593, + 0.004727098159492016, + 0.006956723518669605, + 0.018793858587741852, + 0.006894576363265514, + 0.0057893055491149426, + 0.006124502047896385, + 0.006831355858594179, + 0.005016770213842392, + 0.014637676998972893, + 0.007116840220987797, + 0.007414796855300665, + 0.010832122527062893, + 0.008460372686386108, + 0.005679714493453503, + 0.009465343318879604, + 0.009554526768624783, + 0.009072943590581417, + 0.0013780709123238921, + 0.0023372897412627935, + 0.0029061418026685715, + 0.0019178679212927818, + 0.0020018324721604586, + 0.0016187564469873905, + 0.0028232319746166468, + 0.003113215323537588, + 0.0022413397673517466, + 0.0027622238267213106, + 0.003976938780397177, + 0.0026946004945784807, + 0.004560258239507675, + 0.004329506773501635, + 0.00491330586373806, + 0.0035397042520344257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.004692020360380411 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028271013870835304 + ], + "opposite_logprob_tokens": [ + -5.877827167510986 + ], + "logprob_reward": -0.0028271013870835304, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028271013870835304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13116270303726196, + 0.14300276339054108, + 0.03573217615485191, + 0.10323944687843323, + 0.2032691240310669, + 0.12794406712055206, + 0.11087170988321304, + 0.07975725084543228, + 0.03015918843448162, + 0.03486156463623047 + ], + "credit_map": [ + 0.011915893293917179, + 0.015022201463580132, + 0.014733717776834965, + 0.01451797690242529, + 0.015937741845846176, + 0.015064531937241554, + 0.02931305766105652, + 0.01465759426355362, + 0.019021671265363693, + 0.009793292731046677, + 0.03999714553356171, + 0.007814384996891022, + 0.00551606947556138, + 0.042403701692819595, + 0.007442235480993986, + 0.011014271527528763, + 0.003259435063228011, + 0.005682901479303837, + 0.004623407032340765, + 0.003097591223195195, + 0.003436456900089979, + 0.004050165880471468, + 0.005852666683495045, + 0.005729552824050188, + 0.014553065411746502, + 0.008596940897405148, + 0.018338670954108238, + 0.00943653006106615, + 0.006634814664721489, + 0.014930032193660736, + 0.008930305950343609, + 0.021819090470671654, + 0.03360998257994652, + 0.058312322944402695, + 0.023053569719195366, + 0.008295498788356781, + 0.009578113444149494, + 0.01678803190588951, + 0.012662879191339016, + 0.04096871614456177, + 0.012272978201508522, + 0.013991844840347767, + 0.011620922945439816, + 0.013121265918016434, + 0.025160405784845352, + 0.015691330656409264, + 0.02559717744588852, + 0.010488145053386688, + 0.011930659413337708, + 0.022751452401280403, + 0.0110032232478261, + 0.008655322715640068, + 0.00921234767884016, + 0.01699657551944256, + 0.00900096446275711, + 0.021321162581443787, + 0.009306710213422775, + 0.008349305018782616, + 0.018088994547724724, + 0.010070991702377796, + 0.007122544571757317, + 0.0066713267005980015, + 0.00700983963906765, + 0.01313753705471754, + 0.0027177545707672834, + 0.003280075965449214, + 0.0035842456854879856, + 0.0029230783693492413, + 0.0038112725596874952, + 0.004523035604506731, + 0.005187398754060268, + 0.00413232808932662, + 0.002961412537842989, + 0.0038696236442774534, + 0.004374382551759481, + 0.0026655893307179213, + 0.006218789145350456, + 0.006576609332114458, + 0.004531839396804571, + 0.0036633198615163565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0028271013870835304 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002493250882253051 + ], + "opposite_logprob_tokens": [ + -6.002493381500244 + ], + "logprob_reward": -0.002493250882253051, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002493250882253051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051065582782030106, + 0.07997683435678482, + 0.03954919055104256, + 0.0705605000257492, + 0.10191000252962112, + 0.14471027255058289, + 0.19831982254981995, + 0.19432178139686584, + 0.052105654031038284, + 0.06748037040233612 + ], + "credit_map": [ + 0.006341859232634306, + 0.009306936524808407, + 0.006579139269888401, + 0.004966406151652336, + 0.0066926134750247, + 0.005957478657364845, + 0.006194370333105326, + 0.005026774015277624, + 0.012329553253948689, + 0.010461474768817425, + 0.016573606058955193, + 0.005705410614609718, + 0.004708544351160526, + 0.01857084222137928, + 0.005260516423732042, + 0.006366885732859373, + 0.004395929630845785, + 0.0030392594635486603, + 0.009161917492747307, + 0.0025749204214662313, + 0.0022613380569964647, + 0.007628089282661676, + 0.004851157311350107, + 0.005636575631797314, + 0.008545054122805595, + 0.005485923960804939, + 0.013822510838508606, + 0.00728223193436861, + 0.005965113174170256, + 0.009645156562328339, + 0.0067789992317557335, + 0.013035507872700691, + 0.025015803053975105, + 0.02964811585843563, + 0.01002372894436121, + 0.005281287245452404, + 0.006552106235176325, + 0.009270084090530872, + 0.009395553730428219, + 0.00672332476824522, + 0.01652309112250805, + 0.03679851070046425, + 0.016464339569211006, + 0.008704712614417076, + 0.009259611368179321, + 0.03847246617078781, + 0.008550628088414669, + 0.009936906397342682, + 0.022290874272584915, + 0.038852378726005554, + 0.018008755519986153, + 0.01901984214782715, + 0.019078725948929787, + 0.03002510964870453, + 0.016459021717309952, + 0.03458510711789131, + 0.023452691733837128, + 0.027675719931721687, + 0.0399090051651001, + 0.023355521261692047, + 0.014209218323230743, + 0.021402351558208466, + 0.017657820135354996, + 0.026659458875656128, + 0.004703739657998085, + 0.005144466646015644, + 0.0061789024621248245, + 0.005253368057310581, + 0.007767853792756796, + 0.0065552289597690105, + 0.010204624384641647, + 0.006297471933066845, + 0.011827947571873665, + 0.006099112797528505, + 0.010829965583980083, + 0.005921951495110989, + 0.006787452846765518, + 0.006293929181993008, + 0.010107354260981083, + 0.009612662717700005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.002493250882253051 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775150567293167 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775150567293167, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775150567293167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056197039783000946, + 0.08957080543041229, + 0.031014591455459595, + 0.03392907977104187, + 0.2324083298444748, + 0.0794919952750206, + 0.18604043126106262, + 0.1773751974105835, + 0.04679622873663902, + 0.06717626750469208 + ], + "credit_map": [ + 0.0070296842604875565, + 0.007477676961570978, + 0.007422660943120718, + 0.005587897729128599, + 0.008023208007216454, + 0.006371851544827223, + 0.007718128617852926, + 0.006565937306731939, + 0.012606941163539886, + 0.006631915923207998, + 0.024122539907693863, + 0.0047078304924070835, + 0.003511708229780197, + 0.02605942264199257, + 0.004616414196789265, + 0.007314031943678856, + 0.0031524195801466703, + 0.002565258415415883, + 0.008862129412591457, + 0.0021404826547950506, + 0.0018577254377305508, + 0.0035530528984963894, + 0.0022553822491317987, + 0.006628141272813082, + 0.0038371507544070482, + 0.002927044639363885, + 0.006945707835257053, + 0.003165564266964793, + 0.003119515487924218, + 0.004120383411645889, + 0.004992937669157982, + 0.004820775240659714, + 0.04104915261268616, + 0.07029973715543747, + 0.02621210739016533, + 0.010293697007000446, + 0.00897462572902441, + 0.01818869635462761, + 0.009260006248950958, + 0.04813030734658241, + 0.012171673588454723, + 0.017393747344613075, + 0.010734162293374538, + 0.006885866168886423, + 0.007428774144500494, + 0.013838358223438263, + 0.00412207143381238, + 0.006917340215295553, + 0.02168671041727066, + 0.0392000712454319, + 0.018525242805480957, + 0.014665583148598671, + 0.016802283003926277, + 0.027197793126106262, + 0.014600935392081738, + 0.033361807465553284, + 0.01889820210635662, + 0.020666467025876045, + 0.034848012030124664, + 0.023277562111616135, + 0.01617250218987465, + 0.01619202084839344, + 0.021238595247268677, + 0.026081832125782967, + 0.0046244049444794655, + 0.006787207908928394, + 0.006586895789951086, + 0.005390202160924673, + 0.0047816988080739975, + 0.006632727570831776, + 0.006080655846744776, + 0.005912432912737131, + 0.006500974763184786, + 0.006209616083651781, + 0.010160643607378006, + 0.006507072132080793, + 0.005885496735572815, + 0.008802087977528572, + 0.015420656651258469, + 0.007689716760069132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.009775150567293167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066960170865059 + ], + "opposite_logprob_tokens": [ + -4.511066913604736 + ], + "logprob_reward": -0.011066960170865059, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066960170865059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07585899531841278, + 0.0830410048365593, + 0.03883354365825653, + 0.042925767600536346, + 0.16177846491336823, + 0.14508026838302612, + 0.0788603350520134, + 0.2594508230686188, + 0.05903760716319084, + 0.05513321980834007 + ], + "credit_map": [ + 0.009799099527299404, + 0.011960998177528381, + 0.00934674497693777, + 0.006301256362348795, + 0.007147521711885929, + 0.009182761423289776, + 0.012332494370639324, + 0.009788116440176964, + 0.0104101886972785, + 0.007448364514857531, + 0.02225005254149437, + 0.004463838879019022, + 0.0038323495537042618, + 0.022524047642946243, + 0.004558613523840904, + 0.007553547155112028, + 0.0036722735967487097, + 0.003585825441405177, + 0.008466541767120361, + 0.0034216928761452436, + 0.002945186570286751, + 0.004427480977028608, + 0.004782630130648613, + 0.007531912066042423, + 0.005158868618309498, + 0.0045042335987091064, + 0.006380322854965925, + 0.004677305463701487, + 0.004095456562936306, + 0.006589256692677736, + 0.0054941801354289055, + 0.006026142742484808, + 0.036204565316438675, + 0.05804857239127159, + 0.013448276557028294, + 0.006375014781951904, + 0.0075040520168840885, + 0.01624751091003418, + 0.004949511494487524, + 0.019000966101884842, + 0.018746480345726013, + 0.037046872079372406, + 0.018377594649791718, + 0.008761509321630001, + 0.008085607551038265, + 0.03617427125573158, + 0.007712694350630045, + 0.010175220668315887, + 0.008150694891810417, + 0.007601101417094469, + 0.009479482658207417, + 0.01327476091682911, + 0.011081898584961891, + 0.006983765866607428, + 0.011180236004292965, + 0.011108394712209702, + 0.029604865238070488, + 0.04806884378194809, + 0.040735967457294464, + 0.02669759839773178, + 0.019163286313414574, + 0.04195467382669449, + 0.02325758896768093, + 0.029968004673719406, + 0.0077144308015704155, + 0.007878132164478302, + 0.0063643865287303925, + 0.00587612995877862, + 0.0049164206720888615, + 0.014876078814268112, + 0.004511340521275997, + 0.006900685839354992, + 0.008689929731190205, + 0.005236819852143526, + 0.009830956347286701, + 0.0041115074418485165, + 0.004490198567509651, + 0.008446888998150826, + 0.005475624930113554, + 0.008851294405758381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.011066960170865059 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with an extreme close-up of a brilliantly colored hummingbird, its emerald green feathers gleaming metallically in the soft golden morning light. Its wings vibrate at a rate too rapid to detect with the naked eye as it hovers in mid-air. It carefully dips its needle-thin beak into the depths of a trumpet-shaped, scarlet flower, its petals still shimmering with morning dew. Then, the camera smoothly pulls back, gradually unfolding this microscopic canvas of life, revealing the flower growing in a vibrant green valley. A few quaint wooden cottages dot the valley floor, covered in lush grass and wildflowers, with plumes of smoke rising from their rooftops. The camera continues to steadily and calmly pull back, the field of view becomes increasingly wider, and finally freezes on a magnificent vista, showing the majestic and rugged mountains surrounding the entire idyllic valley. The outlines of the peaks are looming in the morning mist. The tones of the whole picture are warm and fresh, creating a tranquil yet shocking visual experience from the delicacy of microscopic life to the magnificence of macro nature.", + "seed": 4482, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_8/index_4328_seed_4482_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.006559581030160189, + "vlm_logprob": -0.006559581030160189, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.006559581030160189, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010358382016420364 + ], + "opposite_logprob_tokens": [ + -4.7603583335876465 + ], + "logprob_reward": -0.010358382016420364, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010358382016420364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20244960486888885, + 0.024462644010782242, + 0.19586652517318726, + 0.1676023006439209, + 0.16467346251010895, + 0.07139384746551514, + 0.03253834694623947, + 0.08280237019062042, + 0.017033414915204048, + 0.04117748141288757 + ], + "credit_map": [ + 0.022067749872803688, + 0.027991576120257378, + 0.031020009890198708, + 0.0339353121817112, + 0.016009904444217682, + 0.03301292657852173, + 0.018993180245161057, + 0.019418930634856224, + 0.004398431163281202, + 0.002907447051256895, + 0.002372096059843898, + 0.003273958805948496, + 0.002232474274933338, + 0.002329666633158922, + 0.0030322610400617123, + 0.003916309680789709, + 0.04382679611444473, + 0.009444402530789375, + 0.07182261347770691, + 0.025434153154492378, + 0.00911413412541151, + 0.011196601204574108, + 0.011273449286818504, + 0.01375438179820776, + 0.02178128995001316, + 0.01420110184699297, + 0.023440241813659668, + 0.016092311590909958, + 0.021290240809321404, + 0.017472803592681885, + 0.03572963550686836, + 0.01759468764066696, + 0.020352967083454132, + 0.014113818295300007, + 0.017080288380384445, + 0.021496044471859932, + 0.0247664637863636, + 0.029169660061597824, + 0.023623287677764893, + 0.014070923440158367, + 0.017470620572566986, + 0.005141234025359154, + 0.006447954569011927, + 0.008234995417296886, + 0.005975747015327215, + 0.005366233643144369, + 0.006922026630491018, + 0.015835033729672432, + 0.0051749274134635925, + 0.0021521374583244324, + 0.004391723778098822, + 0.0039692893624305725, + 0.0029925955459475517, + 0.0036311408039182425, + 0.0058876690454781055, + 0.004338864237070084, + 0.026142867282032967, + 0.0057184831239283085, + 0.007384900003671646, + 0.0032286238856613636, + 0.023133769631385803, + 0.00493417214602232, + 0.005427437834441662, + 0.006832115352153778, + 0.003292741486802697, + 0.0011701566400006413, + 0.002372757298871875, + 0.0015861941501498222, + 0.0040487563237547874, + 0.0014845926780253649, + 0.0016038625035434961, + 0.0014743533683940768, + 0.013312466442584991, + 0.002873125718906522, + 0.0025859444867819548, + 0.001718236249871552, + 0.014876277185976505, + 0.0026729789096862078, + 0.0016591488383710384, + 0.0014793011359870434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.010358382016420364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746016908437014 + ], + "opposite_logprob_tokens": [ + -5.0067458152771 + ], + "logprob_reward": -0.006746016908437014, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746016908437014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3544129729270935, + 0.04727451130747795, + 0.2608606815338135, + 0.08833180367946625, + 0.06711495667695999, + 0.060661882162094116, + 0.020115753635764122, + 0.0567183680832386, + 0.015055647119879723, + 0.029453402385115623 + ], + "credit_map": [ + 0.035883065313100815, + 0.0378335677087307, + 0.06442383676767349, + 0.045378413051366806, + 0.03119196556508541, + 0.06982948631048203, + 0.03798135370016098, + 0.031891290098428726, + 0.00568681675940752, + 0.002977428724989295, + 0.016442133113741875, + 0.004750149790197611, + 0.003292586887255311, + 0.0025385504122823477, + 0.002263575792312622, + 0.009323270060122013, + 0.06779999285936356, + 0.008080539293587208, + 0.11400381475687027, + 0.02528436854481697, + 0.006285596638917923, + 0.013060138560831547, + 0.006552793551236391, + 0.019793443381786346, + 0.010913866572082043, + 0.008224391378462315, + 0.00829368457198143, + 0.009624475613236427, + 0.011125721968710423, + 0.012597636319696903, + 0.018442919477820396, + 0.009109113365411758, + 0.007814803160727024, + 0.00800943560898304, + 0.004988629370927811, + 0.00563375698402524, + 0.007444627583026886, + 0.008231449872255325, + 0.013888025656342506, + 0.011104228906333447, + 0.006610178854316473, + 0.005109288729727268, + 0.006130173336714506, + 0.00784282200038433, + 0.009377807378768921, + 0.007282051257789135, + 0.009709808975458145, + 0.00859974604099989, + 0.004304786212742329, + 0.001941594877280295, + 0.00245214207097888, + 0.002371216658502817, + 0.0018505442421883345, + 0.0018575498834252357, + 0.0026658810675144196, + 0.00267203850671649, + 0.01835024170577526, + 0.004301092121750116, + 0.004795465152710676, + 0.0016255209920927882, + 0.019532879814505577, + 0.0037297371309250593, + 0.00228864885866642, + 0.002094781259074807, + 0.0037230844609439373, + 0.0008225311175920069, + 0.0012935999548062682, + 0.0012566326186060905, + 0.0021340276580303907, + 0.0014662024332210422, + 0.0023489228915423155, + 0.002010646276175976, + 0.009496652521193027, + 0.001964867115020752, + 0.001823229598812759, + 0.0013285779859870672, + 0.00957261398434639, + 0.001948405522853136, + 0.0017519148532301188, + 0.001567141734994948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.006746016908437014 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013397298753261566 + ], + "opposite_logprob_tokens": [ + -6.626339912414551 + ], + "logprob_reward": -0.0013397298753261566, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013397298753261566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0943043902516365, + 0.06211475282907486, + 0.23846010863780975, + 0.06404824554920197, + 0.13770872354507446, + 0.08461184799671173, + 0.020590418949723244, + 0.17302250862121582, + 0.07331439852714539, + 0.05182459205389023 + ], + "credit_map": [ + 0.0124670946970582, + 0.01795409433543682, + 0.01406578253954649, + 0.009059424512088299, + 0.011539234779775143, + 0.00989878922700882, + 0.009876306168735027, + 0.009443673305213451, + 0.007372822146862745, + 0.003924889490008354, + 0.02093847468495369, + 0.005811898969113827, + 0.003444023197516799, + 0.004723847843706608, + 0.003365378128364682, + 0.012533421628177166, + 0.06250559538602829, + 0.010517564602196217, + 0.10048002004623413, + 0.024291832000017166, + 0.005912325344979763, + 0.013018597848713398, + 0.007219263818114996, + 0.01451493613421917, + 0.012171735987067223, + 0.004819741938263178, + 0.012748626060783863, + 0.007596571464091539, + 0.005531930830329657, + 0.009282322600483894, + 0.006003825459629297, + 0.0058934991247951984, + 0.017965160310268402, + 0.015518519096076488, + 0.012612611055374146, + 0.015699148178100586, + 0.016251543536782265, + 0.025117438286542892, + 0.015837082639336586, + 0.018707217648625374, + 0.015308000147342682, + 0.0069823674857616425, + 0.009739870205521584, + 0.010923742316663265, + 0.007368655875325203, + 0.006663248874247074, + 0.01358950324356556, + 0.014036469161510468, + 0.0029289238154888153, + 0.0021818147506564856, + 0.0035493853501975536, + 0.0024879262782633305, + 0.0018212187569588423, + 0.002538904547691345, + 0.002656353637576103, + 0.00242589320987463, + 0.056480687111616135, + 0.012997403740882874, + 0.01465595979243517, + 0.005338861141353846, + 0.058861225843429565, + 0.011257614940404892, + 0.00726583506911993, + 0.006164934020489454, + 0.0226772278547287, + 0.00470790546387434, + 0.005515187978744507, + 0.00411669397726655, + 0.02197037637233734, + 0.004628562834113836, + 0.0059958468191325665, + 0.003702605841681361, + 0.0165072213858366, + 0.0036002451088279486, + 0.003119288245216012, + 0.0020593840163201094, + 0.01780291274189949, + 0.003377882530912757, + 0.0030369197484105825, + 0.0023207413032650948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.0013397298753261566 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730860564857721 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006730860564857721, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730860564857721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1205308735370636, + 0.026030052453279495, + 0.20307840406894684, + 0.07115470618009567, + 0.061728302389383316, + 0.06749579310417175, + 0.03308078646659851, + 0.16057126224040985, + 0.13868340849876404, + 0.11764637380838394 + ], + "credit_map": [ + 0.014515053480863571, + 0.010733144357800484, + 0.017607644200325012, + 0.019501442089676857, + 0.010462360456585884, + 0.02642422914505005, + 0.00977770984172821, + 0.011509289033710957, + 0.0038154413923621178, + 0.0025070684496313334, + 0.0032986567821353674, + 0.0038012892473489046, + 0.0019030100665986538, + 0.002258888678625226, + 0.00273294560611248, + 0.005712753161787987, + 0.05148783698678017, + 0.007467981893569231, + 0.08562473952770233, + 0.020480036735534668, + 0.00597800686955452, + 0.012042981572449207, + 0.006292499136179686, + 0.013704316690564156, + 0.011612579226493835, + 0.007434926927089691, + 0.012274716049432755, + 0.006259898655116558, + 0.007289942819625139, + 0.00937163457274437, + 0.008703134953975677, + 0.008207873441278934, + 0.007638736627995968, + 0.006458118557929993, + 0.006398322526365519, + 0.008272572420537472, + 0.007095919921994209, + 0.006655518896877766, + 0.008371710777282715, + 0.010837404988706112, + 0.009053963236510754, + 0.006622693967074156, + 0.007307602558284998, + 0.00790388137102127, + 0.010093051008880138, + 0.007401496637612581, + 0.011149151250720024, + 0.007963950745761395, + 0.005485893227159977, + 0.0031882349867373705, + 0.00427386024966836, + 0.0031249835155904293, + 0.003586627310141921, + 0.003141508437693119, + 0.004027052316814661, + 0.006252627354115248, + 0.05333840474486351, + 0.011460823938250542, + 0.013896819204092026, + 0.0049542756751179695, + 0.05072615295648575, + 0.009650158695876598, + 0.007917046546936035, + 0.008627589792013168, + 0.04578755050897598, + 0.008955423720180988, + 0.010871358215808868, + 0.006837901659309864, + 0.04655810445547104, + 0.00839968677610159, + 0.005445414688438177, + 0.005827972665429115, + 0.0375099703669548, + 0.007996799424290657, + 0.007320455741137266, + 0.006199520546942949, + 0.04091303423047066, + 0.007529753725975752, + 0.006168546620756388, + 0.004008289892226458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.006730860564857721 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622915785759687 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622915785759687, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622915785759687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29223546385765076, + 0.030319083482027054, + 0.24730338156223297, + 0.16019229590892792, + 0.04786832630634308, + 0.04668433219194412, + 0.03204591944813728, + 0.05092531070113182, + 0.07761280983686447, + 0.014813068322837353 + ], + "credit_map": [ + 0.030968356877565384, + 0.031290192157030106, + 0.048940397799015045, + 0.03569776564836502, + 0.02933487482368946, + 0.0636347234249115, + 0.022548502311110497, + 0.029820669442415237, + 0.005307409912347794, + 0.003448381321504712, + 0.005552974063903093, + 0.004029312636703253, + 0.002473062602803111, + 0.004012593999505043, + 0.0019111467991024256, + 0.0035842005163431168, + 0.06071050837635994, + 0.011048953048884869, + 0.10187779366970062, + 0.02383498288691044, + 0.008112567476928234, + 0.014921171590685844, + 0.009582001715898514, + 0.017215417698025703, + 0.022555338218808174, + 0.021999232470989227, + 0.017249008640646935, + 0.011820043437182903, + 0.01854736916720867, + 0.03044978529214859, + 0.01705184206366539, + 0.020519673824310303, + 0.006800968665629625, + 0.006137087941169739, + 0.005120973102748394, + 0.005184936802834272, + 0.00627680029720068, + 0.010354122146964073, + 0.003908475395292044, + 0.004084964282810688, + 0.005765275564044714, + 0.004918649792671204, + 0.004625769332051277, + 0.005275780335068703, + 0.005537017714232206, + 0.006081241182982922, + 0.0069493744522333145, + 0.007531226612627506, + 0.00415436364710331, + 0.0029029499273747206, + 0.0045753647573292255, + 0.003860665950924158, + 0.0038871734868735075, + 0.0044725555926561356, + 0.0037482057232409716, + 0.004444639198482037, + 0.01523712370544672, + 0.00428028404712677, + 0.004699643701314926, + 0.002442340599372983, + 0.01044408604502678, + 0.0031428274232894182, + 0.004403961822390556, + 0.0062750414945185184, + 0.0238509364426136, + 0.005384733900427818, + 0.005755788646638393, + 0.0026707828510552645, + 0.02722553350031376, + 0.005032759625464678, + 0.0040764654986560345, + 0.0036158121656626463, + 0.0019486353266984224, + 0.0016061272472143173, + 0.0016463927458971739, + 0.0020427871495485306, + 0.0017027456779032946, + 0.0013108309358358383, + 0.0027042038273066282, + 0.0018513441318646073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.007622915785759687 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The shot begins with an extreme close-up of a brilliantly colored hummingbird, its emerald green feathers gleaming metallically in the soft golden morning light. Its wings vibrate at a rate too rapid to detect with the naked eye as it hovers in mid-air. It carefully dips its needle-thin beak into the depths of a trumpet-shaped, scarlet flower, its petals still shimmering with morning dew. Then, the camera smoothly pulls back, gradually unfolding this microscopic canvas of life, revealing the flower growing in a vibrant green valley. A few quaint wooden cottages dot the valley floor, covered in lush grass and wildflowers, with plumes of smoke rising from their rooftops. The camera continues to steadily and calmly pull back, the field of view becomes increasingly wider, and finally freezes on a magnificent vista, showing the majestic and rugged mountains surrounding the entire idyllic valley. The outlines of the peaks are looming in the morning mist. The tones of the whole picture are warm and fresh, creating a tranquil yet shocking visual experience from the delicacy of microscopic life to the magnificence of macro nature.", + "seed": 4483, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_8/index_4328_seed_4483_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0346668966114521, + "vlm_logprob": -0.0346668966114521, + "vlm_yes_no_margin": 4.399999976158142, + "vlm_token_credit": -0.0346668966114521, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14522871375083923 + ], + "opposite_logprob_tokens": [ + -2.020228624343872 + ], + "logprob_reward": -0.14522871375083923, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14522871375083923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11755399405956268, + 0.03013320453464985, + 0.02892909198999405, + 0.05509622022509575, + 0.03493061289191246, + 0.032603658735752106, + 0.08307648450136185, + 0.10382440686225891, + 0.4958176910877228, + 0.01803460344672203 + ], + "credit_map": [ + 0.015947680920362473, + 0.012572520412504673, + 0.014060069806873798, + 0.014772572554647923, + 0.011109027080237865, + 0.014937776140868664, + 0.020428378134965897, + 0.013725967146456242, + 0.004843650385737419, + 0.0021505996119230986, + 0.005135232117027044, + 0.004642108920961618, + 0.00189334771130234, + 0.0036817092914134264, + 0.0038076566997915506, + 0.0039789010770618916, + 0.0050042616203427315, + 0.0033273568842560053, + 0.003939986694604158, + 0.004240453243255615, + 0.001695391722023487, + 0.002680425764992833, + 0.0034785415045917034, + 0.004562673158943653, + 0.010678286664187908, + 0.006649848073720932, + 0.006961925886571407, + 0.006428784225136042, + 0.004713055677711964, + 0.005901069846004248, + 0.005351954605430365, + 0.008411291986703873, + 0.004117562435567379, + 0.006933798547834158, + 0.0035164193250238895, + 0.003899890463799238, + 0.004241781774908304, + 0.0029390964191406965, + 0.00471872091293335, + 0.0045633441768586636, + 0.004784791264683008, + 0.003534244140610099, + 0.004435149487107992, + 0.004082569386810064, + 0.00256342813372612, + 0.0033647269010543823, + 0.00488077849149704, + 0.004957971628755331, + 0.007190925069153309, + 0.008287428878247738, + 0.012427616864442825, + 0.0105857253074646, + 0.011138278990983963, + 0.005633580032736063, + 0.01618785411119461, + 0.011625074781477451, + 0.006434660404920578, + 0.006823695730417967, + 0.032726891338825226, + 0.006013526115566492, + 0.0056757633574306965, + 0.004032350145280361, + 0.00854268204420805, + 0.03357484191656113, + 0.0881185531616211, + 0.09417539089918137, + 0.036043211817741394, + 0.051965367048978806, + 0.0952066034078598, + 0.036031972616910934, + 0.05368509888648987, + 0.040591511875391006, + 0.001760227489285171, + 0.0032292241230607033, + 0.0023715710267424583, + 0.002566373208537698, + 0.0023575080558657646, + 0.0017086993902921677, + 0.002090789144858718, + 0.0019502119394019246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.14522871375083923 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653522469103336 + ], + "opposite_logprob_tokens": [ + -4.75865364074707 + ], + "logprob_reward": -0.008653522469103336, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653522469103336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16832640767097473, + 0.25950106978416443, + 0.04387904703617096, + 0.06465694308280945, + 0.03626956045627594, + 0.044996969401836395, + 0.14803805947303772, + 0.040369074791669846, + 0.12499213963747025, + 0.06897074729204178 + ], + "credit_map": [ + 0.015551234595477581, + 0.014852960593998432, + 0.01766396127641201, + 0.016826841980218887, + 0.019235290586948395, + 0.02208656258881092, + 0.038082197308540344, + 0.02402736432850361, + 0.026165269315242767, + 0.02068956010043621, + 0.0298377126455307, + 0.03230922296643257, + 0.030447356402873993, + 0.047174710780382156, + 0.039807502180337906, + 0.03306974470615387, + 0.00582386227324605, + 0.003944710362702608, + 0.006295878905802965, + 0.006143953651189804, + 0.004388398956507444, + 0.004276581108570099, + 0.006698360666632652, + 0.006307302974164486, + 0.010243959724903107, + 0.0069551956839859486, + 0.008977743797004223, + 0.007537191268056631, + 0.006602054927498102, + 0.006401873659342527, + 0.00765106501057744, + 0.010287860408425331, + 0.0027462567668408155, + 0.005183166824281216, + 0.004611968994140625, + 0.004659795202314854, + 0.0054582334123551846, + 0.002582014072686434, + 0.005734405945986509, + 0.005293719936162233, + 0.003949781879782677, + 0.005596979055553675, + 0.005402437876909971, + 0.006478630937635899, + 0.005879782140254974, + 0.0031433941330760717, + 0.007524074520915747, + 0.007021888624876738, + 0.016897758468985558, + 0.013313932344317436, + 0.02473963424563408, + 0.023787418380379677, + 0.014267763122916222, + 0.011772653087973595, + 0.023908784613013268, + 0.019350094720721245, + 0.00390741927549243, + 0.005607823375612497, + 0.004928907845169306, + 0.005538949277251959, + 0.005794604774564505, + 0.005615617148578167, + 0.004546848591417074, + 0.004428906831890345, + 0.019358063116669655, + 0.025669271126389503, + 0.011487927287817001, + 0.014818059280514717, + 0.013408374972641468, + 0.009814193472266197, + 0.01632136106491089, + 0.014114885590970516, + 0.005182565189898014, + 0.007732993923127651, + 0.007598023861646652, + 0.011219668202102184, + 0.008391267620027065, + 0.005774208810180426, + 0.009582224301993847, + 0.013489795848727226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.008653522469103336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003196846228092909 + ], + "opposite_logprob_tokens": [ + -5.753196716308594 + ], + "logprob_reward": -0.003196846228092909, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003196846228092909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26964956521987915, + 0.04904070124030113, + 0.028786705806851387, + 0.05669005215167999, + 0.04234718158841133, + 0.05719654634594917, + 0.05143455043435097, + 0.07543234527111053, + 0.30938720703125, + 0.060035157948732376 + ], + "credit_map": [ + 0.03145543113350868, + 0.02195160835981369, + 0.031834352761507034, + 0.03137264773249626, + 0.03121584840118885, + 0.0340687595307827, + 0.0537276491522789, + 0.0340232215821743, + 0.005831544287502766, + 0.005770914256572723, + 0.007221983280032873, + 0.0057780202478170395, + 0.004423555918037891, + 0.007397737354040146, + 0.006570110563188791, + 0.006046832539141178, + 0.0034995886962860823, + 0.0029514043126255274, + 0.004018439911305904, + 0.0026837477926164865, + 0.0037066908553242683, + 0.004363516811281443, + 0.003632593434303999, + 0.003930721897631884, + 0.011474141851067543, + 0.005502272397279739, + 0.007499117404222488, + 0.005653190426528454, + 0.005113393068313599, + 0.006391771137714386, + 0.00646935636177659, + 0.008586804382503033, + 0.00546398526057601, + 0.006284510716795921, + 0.006299552507698536, + 0.0055591450072824955, + 0.0033177968580275774, + 0.0039373161271214485, + 0.004698737058788538, + 0.006786135490983725, + 0.008659865707159042, + 0.009732111357152462, + 0.008708602748811245, + 0.007238711230456829, + 0.0043397205881774426, + 0.004907888825982809, + 0.0062979357317090034, + 0.007311703637242317, + 0.005869628395885229, + 0.005131517071276903, + 0.0073504094034433365, + 0.00695077795535326, + 0.007031944114714861, + 0.004668862093240023, + 0.007782563101500273, + 0.006648838985711336, + 0.006358711048960686, + 0.009398220106959343, + 0.010863864794373512, + 0.008786276914179325, + 0.007767938077449799, + 0.013911825604736805, + 0.009724950417876244, + 0.00862054992467165, + 0.05734090507030487, + 0.06313012540340424, + 0.023778481408953667, + 0.03190382570028305, + 0.05203236639499664, + 0.01908799074590206, + 0.03512680158019066, + 0.026986708864569664, + 0.005991549231112003, + 0.007182711269706488, + 0.008853419683873653, + 0.0085114361718297, + 0.003736308077350259, + 0.004381278995424509, + 0.009944985620677471, + 0.011433466337621212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.003196846228092909 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634022437036037 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008634022437036037, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634022437036037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.300138384103775, + 0.03418036177754402, + 0.028181545436382294, + 0.03305554762482643, + 0.023155346512794495, + 0.03946830332279205, + 0.037229545414447784, + 0.19898982346057892, + 0.16458213329315186, + 0.1410190463066101 + ], + "credit_map": [ + 0.030582427978515625, + 0.026175161823630333, + 0.03225089982151985, + 0.0351482518017292, + 0.035943545401096344, + 0.03619435057044029, + 0.06234200671315193, + 0.0415017306804657, + 0.0035461706575006247, + 0.004073680844157934, + 0.005303722806274891, + 0.005048153456300497, + 0.0026243554893881083, + 0.004010516218841076, + 0.004973767790943384, + 0.004599996376782656, + 0.003254158888012171, + 0.0023625020403414965, + 0.003968722652643919, + 0.0038821855559945107, + 0.0033151607494801283, + 0.004237249959260225, + 0.003979049157351255, + 0.003182516200467944, + 0.005343761760741472, + 0.004030727781355381, + 0.0046745967119932175, + 0.004533473402261734, + 0.002825819654390216, + 0.0026162201538681984, + 0.004311086144298315, + 0.0047198631800711155, + 0.0038491387385874987, + 0.0022851061075925827, + 0.0033427865710109472, + 0.003057728288695216, + 0.0019883320201188326, + 0.002568479161709547, + 0.0026765207294374704, + 0.0033872558269649744, + 0.004826043266803026, + 0.006991397123783827, + 0.004485835786908865, + 0.005474293604493141, + 0.004164762329310179, + 0.00407034158706665, + 0.003876355243846774, + 0.00557927368208766, + 0.00402388721704483, + 0.0037594905588775873, + 0.00469194957986474, + 0.0068165031261742115, + 0.003888525767251849, + 0.0032461693044751883, + 0.006078121718019247, + 0.0047248960472643375, + 0.013172580860555172, + 0.012349107302725315, + 0.053449597209692, + 0.013951714150607586, + 0.00926228053867817, + 0.020064763724803925, + 0.01689320057630539, + 0.05984656885266304, + 0.017051272094249725, + 0.02457532100379467, + 0.012216038070619106, + 0.013069498352706432, + 0.033631641417741776, + 0.02332640439271927, + 0.022883575409650803, + 0.01782839372754097, + 0.013170918449759483, + 0.01449589990079403, + 0.01869581826031208, + 0.02056066133081913, + 0.013536754995584488, + 0.015426402911543846, + 0.02089899405837059, + 0.024233592674136162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.008634022437036037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video an extreme close-up of a hummingbird with emerald green feathers at the beginning?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hummingbird hovering in mid-air while dipping its beak into a scarlet trumpet-shaped flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera smoothly pull back from the close-up to reveal a green valley with wooden cottages and smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the final shot freeze on a wide vista showing rugged mountains surrounding the valley in morning mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the overall lighting warm and golden, consistent with a morning setting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007621378172188997 + ], + "opposite_logprob_tokens": [ + -4.8826212882995605 + ], + "logprob_reward": -0.007621378172188997, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007621378172188997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17754313349723816, + 0.053978100419044495, + 0.02148585580289364, + 0.01674521528184414, + 0.01634438894689083, + 0.03123575635254383, + 0.04476523771882057, + 0.055544957518577576, + 0.060998719185590744, + 0.5213586091995239 + ], + "credit_map": [ + 0.022644083946943283, + 0.01805325783789158, + 0.02185109071433544, + 0.020652079954743385, + 0.019929032772779465, + 0.022507427260279655, + 0.029371511191129684, + 0.02253466099500656, + 0.006424789782613516, + 0.0053651705384254456, + 0.005498830229043961, + 0.00650949077680707, + 0.00529387267306447, + 0.008724741637706757, + 0.008103310130536556, + 0.008057896047830582, + 0.003379160538315773, + 0.002207197481766343, + 0.0022915464360266924, + 0.0027728057466447353, + 0.0021588795352727175, + 0.0026520777028054, + 0.0025267042219638824, + 0.003497485537081957, + 0.002895976183936, + 0.001903593190945685, + 0.0023598605766892433, + 0.002026488771662116, + 0.001326231868006289, + 0.001777442405000329, + 0.0020740048494189978, + 0.002381617436185479, + 0.0021534119732677937, + 0.0017398626077920198, + 0.0017778496257960796, + 0.002540570916607976, + 0.0014337219763547182, + 0.0018780793761834502, + 0.002303178422152996, + 0.0025177146308124065, + 0.00652064336463809, + 0.003907512407749891, + 0.003539471887052059, + 0.003702121088281274, + 0.0023869797587394714, + 0.0038865988608449697, + 0.003261188743636012, + 0.004031240474432707, + 0.003540746169164777, + 0.0040332549251616, + 0.007163273170590401, + 0.007706586737185717, + 0.0050673214718699455, + 0.004204597324132919, + 0.0073226322419941425, + 0.005726823583245277, + 0.005622994154691696, + 0.0041568526066839695, + 0.010726244188845158, + 0.004763377830386162, + 0.004083886742591858, + 0.009520945139229298, + 0.005382479168474674, + 0.011288176290690899, + 0.006335305515676737, + 0.006990313529968262, + 0.007684241980314255, + 0.010433162562549114, + 0.005233443342149258, + 0.008041322231292725, + 0.007036725990474224, + 0.009244206361472607, + 0.04270200803875923, + 0.07186750322580338, + 0.08548574894666672, + 0.10920868068933487, + 0.05892472714185715, + 0.03491513058543205, + 0.06347226351499557, + 0.054782554507255554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.007621378172188997 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_9/index_tensor([9289, 9289])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_9/index_tensor([9289, 9289])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5a76b2dc9686a3c499c4fea8d7d79eb2b814af8c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_9/index_tensor([9289, 9289])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "still, crystal-clear, high-definition close-up captures a scene rendered in the lush, hand-drawn style of '90s Disney animation. A vibrant red-eyed tree frog, its smooth green skin and bright orange feet like suction cups, attempts to hide. Its body lies almost completely flat on the surface of a large, dew-covered tropical leaf, its form defined by the clean, bold outlines characteristic of the animation style. It peers cautiously from behind a dense, emerald green stem, its enormous, perfect scarlet eyes wide in a look of comical panic as it holds its breath and remains silent. The leaf itself is a lifelike masterpiece, with prominent veins and glistening, transparent dewdrops clinging to its waxy surface. Dappled sunlight filtering through the dense rainforest canopy creates warm golden light and soft shadows across the frog and its leafy hideout, illuminating every smooth, vivid detail. Dominated by rich, saturated greens, the scene is punctuated by the vibrant orange and scarlet of the frogs, creating a whimsically charming atmosphere that blends a nostalgic hand-drawn aesthetic with modern high-definition quality.", + "seed": 1693, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_9/index_9289_seed_1693_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1451697576791048, + "vlm_logprob": -0.1451697576791048, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.1451697576791048, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11460436880588531 + ], + "opposite_logprob_tokens": [ + -2.2396044731140137 + ], + "logprob_reward": -0.11460436880588531, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11460436880588531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07092572003602982, + 0.046954426914453506, + 0.03511274605989456, + 0.05150246620178223, + 0.5482630729675293, + 0.11910059303045273, + 0.022202324122190475, + 0.022129978984594345, + 0.016010763123631477, + 0.06779795140028 + ], + "credit_map": [ + 0.007546498905867338, + 0.008194071240723133, + 0.009463715367019176, + 0.010367230512201786, + 0.010059293359518051, + 0.011315345764160156, + 0.006934565957635641, + 0.007044998463243246, + 0.0037753493525087833, + 0.0017765129450708628, + 0.002781403250992298, + 0.0037101779598742723, + 0.013872085139155388, + 0.01646381802856922, + 0.002230643527582288, + 0.002344435779377818, + 0.0039041980635374784, + 0.003812534036114812, + 0.0041541289538145065, + 0.005625673569738865, + 0.0036408170126378536, + 0.005706481169909239, + 0.003921049647033215, + 0.004347864538431168, + 0.005315882153809071, + 0.004889457020908594, + 0.007144587114453316, + 0.010316497646272182, + 0.0047527579590678215, + 0.006445719860494137, + 0.007764054462313652, + 0.004873506724834442, + 0.03851199895143509, + 0.023738401010632515, + 0.013286384753882885, + 0.018604211509227753, + 0.19624969363212585, + 0.21992239356040955, + 0.014125420711934566, + 0.023824559524655342, + 0.009614651091396809, + 0.009416413493454456, + 0.0056856716983020306, + 0.008983956649899483, + 0.030622608959674835, + 0.03613998740911484, + 0.00890412274748087, + 0.00973318237811327, + 0.0022533670999109745, + 0.002452004235237837, + 0.0020533271599560976, + 0.003879052586853504, + 0.00359924859367311, + 0.0033545487094670534, + 0.0021476370748132467, + 0.002463138662278652, + 0.0019947560504078865, + 0.003591718152165413, + 0.0026065653655678034, + 0.0038088702131062746, + 0.0031365889590233564, + 0.0022873191628605127, + 0.002810704754665494, + 0.001893457374535501, + 0.0015354277566075325, + 0.002293383004143834, + 0.0013578776270151138, + 0.0029335611034184694, + 0.0028034185525029898, + 0.0017653948161751032, + 0.001998149324208498, + 0.00132354989182204, + 0.005351508501917124, + 0.0032881011720746756, + 0.00366857647895813, + 0.005037004593759775, + 0.018189413473010063, + 0.025180600583553314, + 0.0033129225485026836, + 0.0037698287051171064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.11460436880588531 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371066227555275 + ], + "opposite_logprob_tokens": [ + -3.4087107181549072 + ], + "logprob_reward": -0.03371066227555275, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371066227555275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10995887219905853, + 0.036796245723962784, + 0.13106736540794373, + 0.03061504289507866, + 0.48791152238845825, + 0.06522494554519653, + 0.02163679338991642, + 0.020615164190530777, + 0.023549119010567665, + 0.07262494415044785 + ], + "credit_map": [ + 0.010391133837401867, + 0.008760972879827023, + 0.010021337307989597, + 0.01350319478660822, + 0.022276271134614944, + 0.024799181148409843, + 0.010575194843113422, + 0.009631588123738766, + 0.0038793832063674927, + 0.003135150531306863, + 0.004420566372573376, + 0.0065721082501113415, + 0.00594788882881403, + 0.005168036092072725, + 0.004399086814373732, + 0.003274025861173868, + 0.015599227510392666, + 0.01327013410627842, + 0.01750136911869049, + 0.01495848037302494, + 0.014865217730402946, + 0.02779518812894821, + 0.012810691259801388, + 0.014267057180404663, + 0.00384174264036119, + 0.0030775221530348063, + 0.0037633716128766537, + 0.00421063182875514, + 0.0029679613653570414, + 0.006103701889514923, + 0.0033002544660121202, + 0.0033498562406748533, + 0.03708683326840401, + 0.01565685123205185, + 0.014416945166885853, + 0.017055410891771317, + 0.16336435079574585, + 0.20324374735355377, + 0.015244175679981709, + 0.021843217313289642, + 0.0070973969995975494, + 0.006673973985016346, + 0.00659578014165163, + 0.006932516116648912, + 0.008132417686283588, + 0.01396334357559681, + 0.006393936928361654, + 0.00943557359278202, + 0.002610909752547741, + 0.0018981961766257882, + 0.002498749177902937, + 0.0034771147184073925, + 0.0021393282804638147, + 0.003740039188414812, + 0.002997548785060644, + 0.0022749064955860376, + 0.002567912684753537, + 0.0018002743599936366, + 0.0029127923771739006, + 0.002745585050433874, + 0.0025356539990752935, + 0.0033603012561798096, + 0.0023380781058222055, + 0.002354564843699336, + 0.0023619476705789566, + 0.0019105783430859447, + 0.0029594283550977707, + 0.004589641932398081, + 0.0027015795931220055, + 0.0033061529975384474, + 0.0030127938371151686, + 0.002706995466724038, + 0.004765328485518694, + 0.0028534147422760725, + 0.003453443292528391, + 0.005132229533046484, + 0.02417774498462677, + 0.02578703872859478, + 0.0030060475692152977, + 0.0034497007727622986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.03371066227555275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809695690870285 + ], + "opposite_logprob_tokens": [ + -3.2880969047546387 + ], + "logprob_reward": -0.03809695690870285, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809695690870285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06257425993680954, + 0.07565420866012573, + 0.02034679800271988, + 0.025348100811243057, + 0.12439657002687454, + 0.5695131421089172, + 0.013280144892632961, + 0.015115279704332352, + 0.013738485053181648, + 0.08003301918506622 + ], + "credit_map": [ + 0.006116240285336971, + 0.005215873010456562, + 0.006471079774200916, + 0.017024686560034752, + 0.007943181321024895, + 0.007345729973167181, + 0.007329634856432676, + 0.005127832759171724, + 0.005511065013706684, + 0.0026904502883553505, + 0.0037146080285310745, + 0.006709054112434387, + 0.021603848785161972, + 0.02673049084842205, + 0.004053026903420687, + 0.00464166235178709, + 0.0025524511002004147, + 0.0015293178148567677, + 0.0020489594899117947, + 0.0048359837383031845, + 0.002025604248046875, + 0.0026515068020671606, + 0.0023656883276998997, + 0.0023372869472950697, + 0.003057115478441119, + 0.0020674339029937983, + 0.002472126856446266, + 0.004279573913663626, + 0.003894778434187174, + 0.004965583793818951, + 0.0020591886714100838, + 0.0025522999931126833, + 0.011344889178872108, + 0.0038526097778230906, + 0.0039859130047261715, + 0.004669344983994961, + 0.036583203822374344, + 0.05191035568714142, + 0.005357804708182812, + 0.00669244397431612, + 0.0433039627969265, + 0.01220428291708231, + 0.012146390043199062, + 0.012930088676512241, + 0.2077743262052536, + 0.24980421364307404, + 0.010235220193862915, + 0.021114688366651535, + 0.0014375915052369237, + 0.001085349591448903, + 0.0021104924380779266, + 0.0018242350779473782, + 0.0017613128293305635, + 0.0023334322031587362, + 0.001315253204666078, + 0.0014124776935204864, + 0.0014342409558594227, + 0.0014948108000680804, + 0.001139733474701643, + 0.0020424332469701767, + 0.0034805522300302982, + 0.002536195795983076, + 0.0014585021417587996, + 0.0015288121066987514, + 0.001606795354746282, + 0.0016135172918438911, + 0.0018834007205441594, + 0.0020793823059648275, + 0.0015622952487319708, + 0.0018451584037393332, + 0.0017441106028854847, + 0.0014038255903869867, + 0.005371888168156147, + 0.0032648644410073757, + 0.0023768788669258356, + 0.0046089389361441135, + 0.027338730171322823, + 0.030762413516640663, + 0.0033032135106623173, + 0.00300609297119081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.03809695690870285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233693718910217 + ], + "opposite_logprob_tokens": [ + -0.8983693718910217 + ], + "logprob_reward": -0.5233693718910217, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233693718910217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036010242998600006, + 0.05008257180452347, + 0.019697433337569237, + 0.02315601520240307, + 0.24272459745407104, + 0.483297199010849, + 0.018266387283802032, + 0.021717660129070282, + 0.0073575847782194614, + 0.09769034385681152 + ], + "credit_map": [ + 0.0028655685018748045, + 0.0027218791656196117, + 0.0029070430900901556, + 0.004012563265860081, + 0.008574198931455612, + 0.010179105214774609, + 0.0022758326958864927, + 0.00247405213303864, + 0.004957959055900574, + 0.0035877921618521214, + 0.004609639756381512, + 0.005320439115166664, + 0.008533060550689697, + 0.014483400620520115, + 0.00420887628570199, + 0.004381401930004358, + 0.0025195800699293613, + 0.001773010822944343, + 0.0017136938404291868, + 0.0040369294583797455, + 0.00209222873672843, + 0.002952262759208679, + 0.002224908210337162, + 0.0023848190903663635, + 0.0029314758721739054, + 0.00179447908885777, + 0.001946166274137795, + 0.004896558355540037, + 0.00386147852987051, + 0.0032638427801430225, + 0.0023453179746866226, + 0.002116695512086153, + 0.017339780926704407, + 0.006616635248064995, + 0.004168323241174221, + 0.006263938266783953, + 0.09060431271791458, + 0.10323339700698853, + 0.004966518841683865, + 0.009531674906611443, + 0.0363960787653923, + 0.010232680477201939, + 0.011554687283933163, + 0.011568119749426842, + 0.1758536547422409, + 0.21104927361011505, + 0.009277455508708954, + 0.0173652321100235, + 0.002266514580696821, + 0.002137962030246854, + 0.0019567934796214104, + 0.00396728003397584, + 0.001675044884905219, + 0.0022140811197459698, + 0.0019411033717915416, + 0.002107608364894986, + 0.0022833594121038914, + 0.003073115600273013, + 0.0021875244565308094, + 0.003483843756839633, + 0.0036401981487870216, + 0.002027790993452072, + 0.0029964004643261433, + 0.0020254270639270544, + 0.000669540895614773, + 0.0006700189551338553, + 0.000969151733443141, + 0.0009745089919306338, + 0.0012087152572348714, + 0.0012250993167981505, + 0.0009507681825198233, + 0.0006897812127135694, + 0.006336148828268051, + 0.0033028132747858763, + 0.0026704345364123583, + 0.004330112133175135, + 0.03420616686344147, + 0.03902323544025421, + 0.0032731336541473866, + 0.0045483047142624855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.5233693718910217 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606742851436138 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606742851436138, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606742851436138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2824341356754303, + 0.15463566780090332, + 0.06163204461336136, + 0.043014634400606155, + 0.04669443890452385, + 0.1757935732603073, + 0.04399283602833748, + 0.03388473764061928, + 0.03162454441189766, + 0.12629343569278717 + ], + "credit_map": [ + 0.026301151141524315, + 0.022373130545020103, + 0.022266190499067307, + 0.061811961233615875, + 0.055339254438877106, + 0.04595078155398369, + 0.026022883132100105, + 0.02236875332891941, + 0.011956754140555859, + 0.0069574336521327496, + 0.007785167545080185, + 0.005541570018976927, + 0.0504540279507637, + 0.058373868465423584, + 0.005580354481935501, + 0.007986483164131641, + 0.007186590228229761, + 0.0062997411005198956, + 0.009795171208679676, + 0.009860211052000523, + 0.005297030787914991, + 0.009618672542273998, + 0.0069792866706848145, + 0.00659533916041255, + 0.004408166278153658, + 0.003679829416796565, + 0.004984625615179539, + 0.00820167362689972, + 0.006764617282897234, + 0.006549476645886898, + 0.005205344874411821, + 0.0032208948396146297, + 0.005458288826048374, + 0.005830681882798672, + 0.005715140141546726, + 0.005025715567171574, + 0.00573289580643177, + 0.010019507259130478, + 0.0035342900082468987, + 0.005377916619181633, + 0.012095874175429344, + 0.005308858118951321, + 0.0051084826700389385, + 0.006475118920207024, + 0.06347259879112244, + 0.07191364467144012, + 0.004407076630741358, + 0.007011918351054192, + 0.004210314713418484, + 0.0028023747727274895, + 0.005876085255295038, + 0.008363993838429451, + 0.0071162693202495575, + 0.006962962448596954, + 0.00457133911550045, + 0.004089495167136192, + 0.003698117332533002, + 0.0024260845966637135, + 0.0035191650968044996, + 0.004969058558344841, + 0.00592289213091135, + 0.006147590931504965, + 0.00407859031111002, + 0.0031232349574565887, + 0.0032688926439732313, + 0.003793668234720826, + 0.004307074472308159, + 0.004920119885355234, + 0.004439680837094784, + 0.0035831599961966276, + 0.003521745791658759, + 0.0037902011536061764, + 0.00883992575109005, + 0.003126834984868765, + 0.0032051601447165012, + 0.005137613043189049, + 0.043939366936683655, + 0.05305885523557663, + 0.0037362417206168175, + 0.005249434150755405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01606742851436138 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, crystal-clear, high-definition close-up captures a scene rendered in the lush, hand-drawn style of '90s Disney animation. A vibrant red-eyed tree frog, its smooth green skin and bright orange feet like suction cups, attempts to hide. Its body lies almost completely flat on the surface of a large, dew-covered tropical leaf, its form defined by the clean, bold outlines characteristic of the animation style. It peers cautiously from behind a dense, emerald green stem, its enormous, perfect scarlet eyes wide in a look of comical panic as it holds its breath and remains silent. The leaf itself is a lifelike masterpiece, with prominent veins and glistening, transparent dewdrops clinging to its waxy surface. Dappled sunlight filtering through the dense rainforest canopy creates warm golden light and soft shadows across the frog and its leafy hideout, illuminating every smooth, vivid detail. Dominated by rich, saturated greens, the scene is punctuated by the vibrant orange and scarlet of the frogs, creating a whimsically charming atmosphere that blends a nostalgic hand-drawn aesthetic with modern high-definition quality.", + "seed": 1694, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_9/index_9289_seed_1694_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08023721873760223, + "vlm_logprob": -0.08023721873760223, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.08023721873760223, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04624101519584656 + ], + "opposite_logprob_tokens": [ + -3.171241044998169 + ], + "logprob_reward": -0.04624101519584656, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04624101519584656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2247246354818344, + 0.06461506336927414, + 0.09164885431528091, + 0.09626135975122452, + 0.10949062556028366, + 0.043520476669073105, + 0.0827745571732521, + 0.13618104159832, + 0.08469101786613464, + 0.06609230488538742 + ], + "credit_map": [ + 0.04083462432026863, + 0.029333414509892464, + 0.015738092362880707, + 0.013979930430650711, + 0.031455233693122864, + 0.049286212772130966, + 0.027189217507839203, + 0.016907915472984314, + 0.010398038662970066, + 0.012659520842134953, + 0.005251443944871426, + 0.009017552249133587, + 0.00937297847121954, + 0.006826654076576233, + 0.0063243890181183815, + 0.004764481447637081, + 0.010737928561866283, + 0.012217211537063122, + 0.009932659566402435, + 0.016039077192544937, + 0.011857434175908566, + 0.011100564152002335, + 0.012276693247258663, + 0.007487288676202297, + 0.013066863641142845, + 0.02177533134818077, + 0.009806899353861809, + 0.010274695232510567, + 0.01241198368370533, + 0.012074781581759453, + 0.009434414096176624, + 0.007416391279548407, + 0.018271097913384438, + 0.02228185348212719, + 0.0077650356106460094, + 0.009843181818723679, + 0.015301670879125595, + 0.016001055017113686, + 0.01255277544260025, + 0.007473956793546677, + 0.0077831679955124855, + 0.007496977224946022, + 0.002868711482733488, + 0.005068766884505749, + 0.0065836915746331215, + 0.005547425244003534, + 0.004142468329519033, + 0.004029270261526108, + 0.014996161684393883, + 0.014377410523593426, + 0.0053844391368329525, + 0.009250764735043049, + 0.009249708615243435, + 0.01051343884319067, + 0.010245849378407001, + 0.00875677727162838, + 0.024231674149632454, + 0.02076566033065319, + 0.009363069199025631, + 0.013965536840260029, + 0.013952280394732952, + 0.02402094192802906, + 0.018952609971165657, + 0.010929282754659653, + 0.012813221663236618, + 0.01598704420030117, + 0.007909270003437996, + 0.006677532568573952, + 0.014119199477136135, + 0.011301252990961075, + 0.007186769507825375, + 0.008696730248630047, + 0.015641065314412117, + 0.009258655831217766, + 0.004513006657361984, + 0.006271973252296448, + 0.009984449483454227, + 0.010792390443384647, + 0.005125503521412611, + 0.00450526550412178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.04624101519584656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059258334338665 + ], + "opposite_logprob_tokens": [ + -3.89559268951416 + ], + "logprob_reward": -0.02059258334338665, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059258334338665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12458251416683197, + 0.13558004796504974, + 0.17281052470207214, + 0.10788217186927795, + 0.08386954665184021, + 0.040558550506830215, + 0.08941156417131424, + 0.06952843070030212, + 0.06495065987110138, + 0.11082594841718674 + ], + "credit_map": [ + 0.02096591331064701, + 0.01718587428331375, + 0.012437568977475166, + 0.010277686640620232, + 0.015138577669858932, + 0.01803402230143547, + 0.012551008723676205, + 0.017991866916418076, + 0.016189010813832283, + 0.016528811305761337, + 0.012548459693789482, + 0.02544999122619629, + 0.0207255557179451, + 0.020470714196562767, + 0.014090538956224918, + 0.009576959535479546, + 0.031240515410900116, + 0.028838777914643288, + 0.014742347411811352, + 0.015920091420412064, + 0.02139955759048462, + 0.029573792591691017, + 0.017201190814375877, + 0.013894247822463512, + 0.0145073551684618, + 0.014872501604259014, + 0.009780541993677616, + 0.009162181988358498, + 0.016587981954216957, + 0.017786923795938492, + 0.011458879336714745, + 0.013725806958973408, + 0.015120928175747395, + 0.01760377734899521, + 0.007115956395864487, + 0.008445528335869312, + 0.011425020173192024, + 0.01068822480738163, + 0.00836870912462473, + 0.005101404618471861, + 0.005790972616523504, + 0.007299947086721659, + 0.003978144843131304, + 0.0039061177521944046, + 0.005403440911322832, + 0.005611351225525141, + 0.004592018201947212, + 0.003976559266448021, + 0.011179308407008648, + 0.01645277999341488, + 0.007471301592886448, + 0.010901040397584438, + 0.013060292229056358, + 0.012892204336822033, + 0.008447444997727871, + 0.009007195942103863, + 0.00931733101606369, + 0.011224955320358276, + 0.0060003893449902534, + 0.008645228110253811, + 0.008199866861104965, + 0.01029240619391203, + 0.008157527074217796, + 0.00769072724506259, + 0.008549017831683159, + 0.009686946868896484, + 0.0059525249525904655, + 0.007750022690743208, + 0.009257609024643898, + 0.008810891769826412, + 0.00788211915642023, + 0.007061527110636234, + 0.018287418410182, + 0.015837980434298515, + 0.010527757927775383, + 0.011788717471063137, + 0.014594019390642643, + 0.014505761675536633, + 0.01306826714426279, + 0.012216025032103062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.02059258334338665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786483831703663 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786483831703663, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786483831703663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1862519383430481, + 0.0830475315451622, + 0.19575649499893188, + 0.13415123522281647, + 0.10644043236970901, + 0.026600968092679977, + 0.11818423122167587, + 0.04979494959115982, + 0.03673629090189934, + 0.06303589046001434 + ], + "credit_map": [ + 0.04153889790177345, + 0.026147421449422836, + 0.012822831980884075, + 0.011635364964604378, + 0.020970921963453293, + 0.03725874423980713, + 0.02142385020852089, + 0.014453905634582043, + 0.009620466269552708, + 0.009683776646852493, + 0.007685964927077293, + 0.011600060388445854, + 0.015169886872172356, + 0.010621592402458191, + 0.008249052800238132, + 0.010416730307042599, + 0.03958999738097191, + 0.027992449700832367, + 0.013449554331600666, + 0.02123899944126606, + 0.030911192297935486, + 0.030835263431072235, + 0.02072407305240631, + 0.011014964431524277, + 0.023083070293068886, + 0.02251957543194294, + 0.014513196423649788, + 0.011044981889426708, + 0.01972731575369835, + 0.021846825256943703, + 0.00995064526796341, + 0.011465623043477535, + 0.017048371955752373, + 0.02210533246397972, + 0.006754026282578707, + 0.011192419566214085, + 0.016244014725089073, + 0.015619616955518723, + 0.010382280685007572, + 0.0070943683385849, + 0.00358921498991549, + 0.004586680326610804, + 0.002752401400357485, + 0.0025233295746147633, + 0.0037375346291810274, + 0.003693843260407448, + 0.0029191216453909874, + 0.0027988431975245476, + 0.014637802727520466, + 0.024112114682793617, + 0.009036670438945293, + 0.014806372113525867, + 0.015528851188719273, + 0.018323365598917007, + 0.012364434078335762, + 0.00937462504953146, + 0.007207710761576891, + 0.007804361637681723, + 0.0035856678150594234, + 0.005499427206814289, + 0.006317734252661467, + 0.00820085033774376, + 0.006545207928866148, + 0.0046339877881109715, + 0.006679223850369453, + 0.005891656968742609, + 0.0028292806819081306, + 0.003848643507808447, + 0.004611556883901358, + 0.006186137441545725, + 0.0032433695159852505, + 0.003446424612775445, + 0.016330944374203682, + 0.008274241350591183, + 0.005273160990327597, + 0.005520611070096493, + 0.00839257426559925, + 0.007778326515108347, + 0.006406762171536684, + 0.0050592729821801186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.009786483831703663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134925663471222 + ], + "opposite_logprob_tokens": [ + -1.3134925365447998 + ], + "logprob_reward": -0.3134925663471222, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134925663471222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0994354858994484, + 0.08458862453699112, + 0.13206373155117035, + 0.10958988964557648, + 0.11781845986843109, + 0.1012912467122078, + 0.045581281185150146, + 0.10662151873111725, + 0.03748435899615288, + 0.1655253916978836 + ], + "credit_map": [ + 0.016735775396227837, + 0.015254940837621689, + 0.006840013898909092, + 0.007400854956358671, + 0.015200206078588963, + 0.01542443037033081, + 0.010657482780516148, + 0.011921782977879047, + 0.009096607565879822, + 0.015128065831959248, + 0.007897518575191498, + 0.017416536808013916, + 0.011697893962264061, + 0.010745558887720108, + 0.007481379900127649, + 0.005125059746205807, + 0.02892765775322914, + 0.020829884335398674, + 0.010755505412817001, + 0.013277968391776085, + 0.014690042473375797, + 0.020219091325998306, + 0.012207873165607452, + 0.011155710555613041, + 0.02226293459534645, + 0.02395692467689514, + 0.009213097393512726, + 0.009340430609881878, + 0.012334724888205528, + 0.0152223976328969, + 0.009556502103805542, + 0.007702877279371023, + 0.016818996518850327, + 0.022685052827000618, + 0.008876495994627476, + 0.010708419606089592, + 0.019743168726563454, + 0.01542066503316164, + 0.013126838952302933, + 0.010438822209835052, + 0.013358678668737411, + 0.019872333854436874, + 0.007493122946470976, + 0.012941312976181507, + 0.016885006800293922, + 0.015050426125526428, + 0.008659213781356812, + 0.007031145505607128, + 0.00608451571315527, + 0.008380820974707603, + 0.0037515058647841215, + 0.006263147573918104, + 0.006323567591607571, + 0.005143464542925358, + 0.00582290580496192, + 0.00381135824136436, + 0.01614430733025074, + 0.016896067187190056, + 0.009168081916868687, + 0.014311626553535461, + 0.01165363471955061, + 0.01486091036349535, + 0.015012414194643497, + 0.008574479259550571, + 0.004425546154379845, + 0.006167745217680931, + 0.0035812649875879288, + 0.0033703246153891087, + 0.007309754844754934, + 0.006795793771743774, + 0.0027991922106593847, + 0.003034737892448902, + 0.027647459879517555, + 0.024387020617723465, + 0.01342870481312275, + 0.014958473853766918, + 0.03524439036846161, + 0.022498304024338722, + 0.015184061601758003, + 0.012176972813904285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.3134925663471222 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073444969952106 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073444969952106, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073444969952106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23874013125896454, + 0.09744241833686829, + 0.18053345382213593, + 0.12111853808164597, + 0.025112954899668694, + 0.014037080109119415, + 0.0500844269990921, + 0.04628552496433258, + 0.017631588503718376, + 0.2090139389038086 + ], + "credit_map": [ + 0.05316125229001045, + 0.04148969426751137, + 0.018676767125725746, + 0.013560122810304165, + 0.027012431994080544, + 0.04769628867506981, + 0.019083674997091293, + 0.018059898167848587, + 0.015987832099199295, + 0.01546495035290718, + 0.008053199388086796, + 0.013399003073573112, + 0.0130071472376585, + 0.012499961070716381, + 0.011440333910286427, + 0.007589997723698616, + 0.029584910720586777, + 0.02583531104028225, + 0.014697178266942501, + 0.02501034550368786, + 0.025765564292669296, + 0.023412294685840607, + 0.02237386628985405, + 0.013853984884917736, + 0.0172751322388649, + 0.02977881021797657, + 0.009299189783632755, + 0.010447841137647629, + 0.017155449837446213, + 0.01801571249961853, + 0.010508459992706776, + 0.008637947961688042, + 0.0038247141055762768, + 0.0035652348306030035, + 0.0021429331973195076, + 0.0029340237379074097, + 0.00395387364551425, + 0.0032288655638694763, + 0.0032513306941837072, + 0.0022119793575257063, + 0.0020021386444568634, + 0.002741242293268442, + 0.0012569044483825564, + 0.001325581455603242, + 0.0021500277798622847, + 0.0017196647822856903, + 0.0015848473412916064, + 0.0012566731311380863, + 0.00822040531784296, + 0.0090888487175107, + 0.0037468220107257366, + 0.004543006420135498, + 0.005949781276285648, + 0.0065870219841599464, + 0.006387900095432997, + 0.005560638848692179, + 0.007725498639047146, + 0.00762599054723978, + 0.0032773660495877266, + 0.00551138911396265, + 0.005132661201059818, + 0.007758807856589556, + 0.005560735706239939, + 0.0036930751521140337, + 0.0024925933685153723, + 0.0024394432548433542, + 0.0015286298003047705, + 0.0014898002846166492, + 0.0027976229321211576, + 0.0025159968063235283, + 0.0020963724236935377, + 0.0022711290512233973, + 0.03540196642279625, + 0.035310957580804825, + 0.016753213480114937, + 0.02067524939775467, + 0.03229144215583801, + 0.03698915243148804, + 0.017248257994651794, + 0.01434369571506977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.011073444969952106 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, crystal-clear, high-definition close-up captures a scene rendered in the lush, hand-drawn style of '90s Disney animation. A vibrant red-eyed tree frog, its smooth green skin and bright orange feet like suction cups, attempts to hide. Its body lies almost completely flat on the surface of a large, dew-covered tropical leaf, its form defined by the clean, bold outlines characteristic of the animation style. It peers cautiously from behind a dense, emerald green stem, its enormous, perfect scarlet eyes wide in a look of comical panic as it holds its breath and remains silent. The leaf itself is a lifelike masterpiece, with prominent veins and glistening, transparent dewdrops clinging to its waxy surface. Dappled sunlight filtering through the dense rainforest canopy creates warm golden light and soft shadows across the frog and its leafy hideout, illuminating every smooth, vivid detail. Dominated by rich, saturated greens, the scene is punctuated by the vibrant orange and scarlet of the frogs, creating a whimsically charming atmosphere that blends a nostalgic hand-drawn aesthetic with modern high-definition quality.", + "seed": 1695, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_9/index_9289_seed_1695_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07579249646514655, + "vlm_logprob": -0.07579249646514655, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.07579249646514655, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0808999165892601 + ], + "opposite_logprob_tokens": [ + -2.580899953842163 + ], + "logprob_reward": -0.0808999165892601, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0808999165892601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18019525706768036, + 0.09913352131843567, + 0.13351018726825714, + 0.07240961492061615, + 0.1973673552274704, + 0.12846946716308594, + 0.07796066999435425, + 0.041829533874988556, + 0.03385310620069504, + 0.035271335393190384 + ], + "credit_map": [ + 0.03270294889807701, + 0.01889886148273945, + 0.012339525856077671, + 0.017299814149737358, + 0.0199833232909441, + 0.02745843306183815, + 0.022740470245480537, + 0.028771892189979553, + 0.023021165281534195, + 0.010342510417103767, + 0.008141647092998028, + 0.010310645215213299, + 0.011304844170808792, + 0.018527358770370483, + 0.006252611055970192, + 0.011232736520469189, + 0.02585405856370926, + 0.013962768949568272, + 0.013866431079804897, + 0.015343143604695797, + 0.01613343507051468, + 0.020449524745345116, + 0.011274680495262146, + 0.016626153141260147, + 0.012563926167786121, + 0.008964580483734608, + 0.007446341682225466, + 0.008921111933887005, + 0.005599668249487877, + 0.0141573715955019, + 0.0062238299287855625, + 0.008532783947885036, + 0.03457155078649521, + 0.028804142028093338, + 0.019602950662374496, + 0.024971989914774895, + 0.013191581703722477, + 0.03917462006211281, + 0.01771232858300209, + 0.01933818683028221, + 0.019960207864642143, + 0.017069833353161812, + 0.01315562054514885, + 0.018462950363755226, + 0.010767174884676933, + 0.023269692435860634, + 0.013605390675365925, + 0.012178593315184116, + 0.014792446047067642, + 0.008993741124868393, + 0.008284717798233032, + 0.008686358109116554, + 0.005364240147173405, + 0.014438574202358723, + 0.007990921847522259, + 0.00940967258065939, + 0.010285167023539543, + 0.008510416373610497, + 0.003235581796616316, + 0.0036861749831587076, + 0.002843480557203293, + 0.006468229461461306, + 0.003477755468338728, + 0.00332272588275373, + 0.0061684874817729, + 0.005728395190089941, + 0.0027713521849364042, + 0.003974889405071735, + 0.0025688426103442907, + 0.0051858630031347275, + 0.0037937737070024014, + 0.0036615037824958563, + 0.005705593153834343, + 0.004854896105825901, + 0.0035675722174346447, + 0.0034160027280449867, + 0.003466374706476927, + 0.00638266047462821, + 0.003863120684400201, + 0.004015113227069378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0808999165892601 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789811834692955 + ], + "opposite_logprob_tokens": [ + -2.5789811611175537 + ], + "logprob_reward": -0.0789811834692955, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789811834692955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09633221477270126, + 0.10523255169391632, + 0.12156357616186142, + 0.10400928556919098, + 0.14203009009361267, + 0.13442054390907288, + 0.1232873722910881, + 0.1159873679280281, + 0.016820715740323067, + 0.04031623899936676 + ], + "credit_map": [ + 0.019229182973504066, + 0.0135377561673522, + 0.009866914711892605, + 0.008818788453936577, + 0.01029665395617485, + 0.013713927939534187, + 0.00875033438205719, + 0.01211866457015276, + 0.021990517154335976, + 0.01176097895950079, + 0.007584413047879934, + 0.012113351374864578, + 0.007677627727389336, + 0.019013792276382446, + 0.007866346277296543, + 0.01722552813589573, + 0.025106526911258698, + 0.019119087606668472, + 0.00922904722392559, + 0.018314672634005547, + 0.008124700747430325, + 0.01967219077050686, + 0.011411448940634727, + 0.01058590691536665, + 0.016585586592555046, + 0.01123806368559599, + 0.010938341729342937, + 0.01850394904613495, + 0.008826866745948792, + 0.014054295606911182, + 0.012047610245645046, + 0.011814572848379612, + 0.02037467062473297, + 0.017850236967206, + 0.013645296916365623, + 0.022164635360240936, + 0.018314791843295097, + 0.018524471670389175, + 0.019367169588804245, + 0.01178883295506239, + 0.032233357429504395, + 0.02571205049753189, + 0.011895958334207535, + 0.012970929965376854, + 0.007981727831065655, + 0.022196508944034576, + 0.01244635321199894, + 0.008983672596514225, + 0.022962013259530067, + 0.012066070921719074, + 0.010201537981629372, + 0.01890450157225132, + 0.009466980583965778, + 0.021473171189427376, + 0.012308655306696892, + 0.01590445451438427, + 0.01607700064778328, + 0.013699266128242016, + 0.012380325235426426, + 0.01809762604534626, + 0.008643651381134987, + 0.016828471794724464, + 0.016917459666728973, + 0.013343583792448044, + 0.0025916548911482096, + 0.002111125970259309, + 0.0019849473610520363, + 0.002437915885820985, + 0.001686768257059157, + 0.0021837472449988127, + 0.001805292209610343, + 0.0020192654337733984, + 0.005075162276625633, + 0.005205197725445032, + 0.004622455686330795, + 0.0050050304271280766, + 0.004995020106434822, + 0.004873224999755621, + 0.005589495413005352, + 0.0049506560899317265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.0789811834692955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062045708298683167 + ], + "opposite_logprob_tokens": [ + -2.8120458126068115 + ], + "logprob_reward": -0.062045708298683167, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062045708298683167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16503119468688965, + 0.12106817960739136, + 0.14033204317092896, + 0.0979880839586258, + 0.08277708292007446, + 0.14490969479084015, + 0.14427457749843597, + 0.05438563600182533, + 0.022428113967180252, + 0.026805365458130836 + ], + "credit_map": [ + 0.03335355222225189, + 0.019670123234391212, + 0.013180851936340332, + 0.011117493733763695, + 0.021762030199170113, + 0.027603141963481903, + 0.01813913695514202, + 0.020204879343509674, + 0.02884829416871071, + 0.012060673907399178, + 0.009652146138250828, + 0.007899186573922634, + 0.011418520472943783, + 0.021179689094424248, + 0.009135022759437561, + 0.02087464928627014, + 0.02719327248632908, + 0.017007026821374893, + 0.01400080882012844, + 0.019851356744766235, + 0.01263787504285574, + 0.01879950240254402, + 0.016244031488895416, + 0.01459816750138998, + 0.01766761764883995, + 0.017986316233873367, + 0.008324984461069107, + 0.014516917988657951, + 0.0070108333602547646, + 0.014761149883270264, + 0.009358919225633144, + 0.008361351676285267, + 0.012754839845001698, + 0.010768250562250614, + 0.009193189442157745, + 0.012835321947932243, + 0.005979133769869804, + 0.013288041576743126, + 0.009719784371554852, + 0.008238526992499828, + 0.03476199135184288, + 0.03071908839046955, + 0.009942726232111454, + 0.014413496479392052, + 0.010188060812652111, + 0.0233208779245615, + 0.010731035843491554, + 0.010832443833351135, + 0.034017935395240784, + 0.025487657636404037, + 0.01286693848669529, + 0.011126945726573467, + 0.008031182922422886, + 0.028536222875118256, + 0.01282300055027008, + 0.011384710669517517, + 0.00666036969050765, + 0.00553926220163703, + 0.004257528576999903, + 0.008632555603981018, + 0.006449471693485975, + 0.00626492640003562, + 0.011412718333303928, + 0.005168807227164507, + 0.00410686107352376, + 0.004009731579571962, + 0.0017414387548342347, + 0.0018575136782601476, + 0.0019612000323832035, + 0.003405983094125986, + 0.0025075653102248907, + 0.0028378234710544348, + 0.005483342334628105, + 0.0049560219049453735, + 0.002439838368445635, + 0.001826271414756775, + 0.0026931872125715017, + 0.004453960340470076, + 0.002054434735327959, + 0.0028983112424612045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.062045708298683167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284272491931915 + ], + "opposite_logprob_tokens": [ + -2.0178427696228027 + ], + "logprob_reward": -0.14284272491931915, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14284272491931915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12536537647247314, + 0.14287760853767395, + 0.1403883397579193, + 0.22293692827224731, + 0.08906218409538269, + 0.11936835199594498, + 0.0574643649160862, + 0.04658897966146469, + 0.028406720608472824, + 0.02754116803407669 + ], + "credit_map": [ + 0.04095815494656563, + 0.015890242531895638, + 0.010082310996949673, + 0.005985674913972616, + 0.010754328221082687, + 0.022790849208831787, + 0.00864436011761427, + 0.010259448550641537, + 0.023364994674921036, + 0.02470124326646328, + 0.0142446868121624, + 0.018879957497119904, + 0.011517598293721676, + 0.01918102614581585, + 0.015563256107270718, + 0.015424845740199089, + 0.028218090534210205, + 0.014376671053469181, + 0.01693444512784481, + 0.014137251302599907, + 0.01333926897495985, + 0.02299225889146328, + 0.015280780382454395, + 0.015109571628272533, + 0.036549124866724014, + 0.02589374966919422, + 0.024937298148870468, + 0.033027298748493195, + 0.020637301728129387, + 0.035197336226701736, + 0.025400390848517418, + 0.021294431760907173, + 0.018037034198641777, + 0.013573000207543373, + 0.006741797085851431, + 0.008800270967185497, + 0.011410726234316826, + 0.01705627515912056, + 0.006893540732562542, + 0.006549547892063856, + 0.01615287736058235, + 0.016556523740291595, + 0.009502389468252659, + 0.016610905528068542, + 0.015345057472586632, + 0.02231105975806713, + 0.011329913511872292, + 0.011559626087546349, + 0.012239081785082817, + 0.006813260726630688, + 0.006189568433910608, + 0.005665705073624849, + 0.00427987240254879, + 0.013987396843731403, + 0.004442567937076092, + 0.0038469096180051565, + 0.010066481307148933, + 0.0078052026219666, + 0.0033568621147423983, + 0.00536052230745554, + 0.0038521920796483755, + 0.00819359254091978, + 0.0038678981363773346, + 0.004086228087544441, + 0.0053536477498710155, + 0.0037151931319385767, + 0.002445430727675557, + 0.0032910602167248726, + 0.00215448928065598, + 0.005334427580237389, + 0.0034661078825592995, + 0.002646365202963352, + 0.00427956273779273, + 0.0037018018774688244, + 0.00304024969227612, + 0.0030639527831226587, + 0.0025732468347996473, + 0.00407132413238287, + 0.00363610009662807, + 0.0031749282497912645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.14284272491931915 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192949049174786 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192949049174786, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192949049174786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1794920414686203, + 0.14534629881381989, + 0.1288454681634903, + 0.07082006335258484, + 0.09772729128599167, + 0.12439123541116714, + 0.09269185364246368, + 0.04482375085353851, + 0.036405861377716064, + 0.0794561356306076 + ], + "credit_map": [ + 0.039318908005952835, + 0.031959619373083115, + 0.017603671178221703, + 0.017310136929154396, + 0.015246337279677391, + 0.027337152510881424, + 0.012575526721775532, + 0.01814068853855133, + 0.03082527406513691, + 0.02401631325483322, + 0.011458426713943481, + 0.018332358449697495, + 0.012891175225377083, + 0.020309366285800934, + 0.012927592732012272, + 0.014585786499083042, + 0.022199394181370735, + 0.016033628955483437, + 0.013881065882742405, + 0.017257580533623695, + 0.011976898647844791, + 0.018114274367690086, + 0.015494444407522678, + 0.013888178393244743, + 0.01260750088840723, + 0.010755477473139763, + 0.007780895102769136, + 0.008864916861057281, + 0.0049169533886015415, + 0.01216631755232811, + 0.005947145167738199, + 0.007780857849866152, + 0.017370158806443214, + 0.013585085980594158, + 0.009131934493780136, + 0.012670348398387432, + 0.006275196094065905, + 0.01966329663991928, + 0.00830415915697813, + 0.010727109387516975, + 0.022821856662631035, + 0.014918508939445019, + 0.011835440061986446, + 0.015714209526777267, + 0.010354136116802692, + 0.024081120267510414, + 0.012340047396719456, + 0.012325920164585114, + 0.017637332901358604, + 0.011128118261694908, + 0.012565074488520622, + 0.009108184836804867, + 0.006225925404578447, + 0.016197167336940765, + 0.008602679707109928, + 0.011227375827729702, + 0.011433576233685017, + 0.008497967384755611, + 0.0034566756803542376, + 0.0036430582404136658, + 0.0036160247400403023, + 0.007178459316492081, + 0.00332074542529881, + 0.0036772447638213634, + 0.0052365646697580814, + 0.005386278498917818, + 0.003414988284930587, + 0.003944327123463154, + 0.004472770262509584, + 0.005110794212669134, + 0.0054859379306435585, + 0.003354201791808009, + 0.014915700070559978, + 0.009336112067103386, + 0.007657639682292938, + 0.0061811357736587524, + 0.008109193295240402, + 0.016503967344760895, + 0.007754788268357515, + 0.008997602388262749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014192949049174786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, crystal-clear, high-definition close-up captures a scene rendered in the lush, hand-drawn style of '90s Disney animation. A vibrant red-eyed tree frog, its smooth green skin and bright orange feet like suction cups, attempts to hide. Its body lies almost completely flat on the surface of a large, dew-covered tropical leaf, its form defined by the clean, bold outlines characteristic of the animation style. It peers cautiously from behind a dense, emerald green stem, its enormous, perfect scarlet eyes wide in a look of comical panic as it holds its breath and remains silent. The leaf itself is a lifelike masterpiece, with prominent veins and glistening, transparent dewdrops clinging to its waxy surface. Dappled sunlight filtering through the dense rainforest canopy creates warm golden light and soft shadows across the frog and its leafy hideout, illuminating every smooth, vivid detail. Dominated by rich, saturated greens, the scene is punctuated by the vibrant orange and scarlet of the frogs, creating a whimsically charming atmosphere that blends a nostalgic hand-drawn aesthetic with modern high-definition quality.", + "seed": 1696, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000064/group_9/index_9289_seed_1696_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16375725101679564, + "vlm_logprob": -0.16375725101679564, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.16375725101679564, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08115242421627045 + ], + "opposite_logprob_tokens": [ + -2.5811524391174316 + ], + "logprob_reward": -0.08115242421627045, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08115242421627045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24524199962615967, + 0.106114462018013, + 0.17847716808319092, + 0.09566515684127808, + 0.0669112503528595, + 0.03759002685546875, + 0.07313713431358337, + 0.10868465155363083, + 0.06530772894620895, + 0.022870443761348724 + ], + "credit_map": [ + 0.034626856446266174, + 0.04535163938999176, + 0.02065783552825451, + 0.017805777490139008, + 0.041111093014478683, + 0.043021779507398605, + 0.017021622508764267, + 0.025645390152931213, + 0.013996430672705173, + 0.017798112705349922, + 0.010727562010288239, + 0.010368164628744125, + 0.013409250415861607, + 0.018541375175118446, + 0.010368525050580502, + 0.010905046947300434, + 0.021519474685192108, + 0.01950603537261486, + 0.027712225914001465, + 0.025034945458173752, + 0.028023844584822655, + 0.024885928258299828, + 0.018508831039071083, + 0.013285891152918339, + 0.017680857330560684, + 0.00918944738805294, + 0.009043795987963676, + 0.010291936807334423, + 0.008466845378279686, + 0.013186411000788212, + 0.01552014984190464, + 0.012285708449780941, + 0.013957184739410877, + 0.006625622045248747, + 0.004275592043995857, + 0.005017062649130821, + 0.00817174930125475, + 0.008615483529865742, + 0.006868651136755943, + 0.013379903510212898, + 0.010505457408726215, + 0.003667625365778804, + 0.00258061895146966, + 0.0022336572874337435, + 0.004733262117952108, + 0.005001249257475138, + 0.002038838341832161, + 0.006829317659139633, + 0.015013642609119415, + 0.00828879326581955, + 0.005947629921138287, + 0.006030835211277008, + 0.00826046708971262, + 0.009514793753623962, + 0.011178847402334213, + 0.008902122266590595, + 0.02416602335870266, + 0.011298594996333122, + 0.006430312525480986, + 0.007101404946297407, + 0.01665372960269451, + 0.01066560484468937, + 0.008743125014007092, + 0.023625856265425682, + 0.01577761583030224, + 0.0059736077673733234, + 0.003329841885715723, + 0.004704176913946867, + 0.00790620967745781, + 0.0071626948192715645, + 0.004829603713005781, + 0.015623977407813072, + 0.004534883890300989, + 0.0020161233842372894, + 0.0020159357227385044, + 0.0024290462024509907, + 0.002226191572844982, + 0.0046376013197004795, + 0.00258596776984632, + 0.0024246934335678816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.08115242421627045 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002881914377213 + ], + "opposite_logprob_tokens": [ + -2.695028781890869 + ], + "logprob_reward": -0.07002881914377213, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002881914377213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29548946022987366, + 0.05157489329576492, + 0.1089668720960617, + 0.15964306890964508, + 0.07009831815958023, + 0.05826402083039284, + 0.04239804670214653, + 0.05229099094867706, + 0.0812436044216156, + 0.08003067225217819 + ], + "credit_map": [ + 0.03762391582131386, + 0.04330127313733101, + 0.0390472337603569, + 0.041065480560064316, + 0.04062451422214508, + 0.049415554851293564, + 0.0205878596752882, + 0.023823631927371025, + 0.007327974308282137, + 0.006683118175715208, + 0.004498184192925692, + 0.005528814159333706, + 0.01118418388068676, + 0.008522236719727516, + 0.0038431554567068815, + 0.0039872280322015285, + 0.014165457338094711, + 0.011809487827122211, + 0.012457776814699173, + 0.01709689013659954, + 0.012900293804705143, + 0.014627257362008095, + 0.014561488293111324, + 0.011348223313689232, + 0.026626523584127426, + 0.01593935489654541, + 0.012319125235080719, + 0.03357385843992233, + 0.015072023496031761, + 0.020059959962964058, + 0.01859590783715248, + 0.017456302419304848, + 0.011023009195923805, + 0.006503795273602009, + 0.005523189436644316, + 0.012318751774728298, + 0.008165561594069004, + 0.008014511317014694, + 0.007571537513285875, + 0.010977965779602528, + 0.008905347436666489, + 0.006548121105879545, + 0.0038355523720383644, + 0.00851320568472147, + 0.008310829289257526, + 0.006594733335077763, + 0.00491879740729928, + 0.010637433268129826, + 0.0124704884365201, + 0.0034892139956355095, + 0.0024837013334035873, + 0.004709554836153984, + 0.0034335616510361433, + 0.003953536972403526, + 0.004843015223741531, + 0.007014979608356953, + 0.012527910061180592, + 0.004442829173058271, + 0.0035312874242663383, + 0.006321354303508997, + 0.005389359779655933, + 0.004385663196444511, + 0.00951410923153162, + 0.0061784773133695126, + 0.019612636417150497, + 0.007812770083546638, + 0.006625905632972717, + 0.009072160348296165, + 0.006114238407462835, + 0.008922065608203411, + 0.009345349855720997, + 0.013738474808633327, + 0.015299002639949322, + 0.006958039943128824, + 0.00719835702329874, + 0.009626339189708233, + 0.007015835028141737, + 0.007025682367384434, + 0.016521602869033813, + 0.010385815054178238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.07002881914377213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201059006154537 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014201059006154537, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201059006154537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1695382297039032, + 0.1022835299372673, + 0.1879144012928009, + 0.08903217315673828, + 0.07933475822210312, + 0.08896328508853912, + 0.049304671585559845, + 0.05508783832192421, + 0.1429138034582138, + 0.035627394914627075 + ], + "credit_map": [ + 0.028791900724172592, + 0.02034955471754074, + 0.02205905318260193, + 0.015139163471758366, + 0.025697728618979454, + 0.024157220497727394, + 0.015506536699831486, + 0.01783706247806549, + 0.012793817557394505, + 0.011332438327372074, + 0.011111103929579258, + 0.01932966709136963, + 0.009526013396680355, + 0.017847031354904175, + 0.012407035566866398, + 0.007936419919133186, + 0.02355450950562954, + 0.02533900924026966, + 0.02465115860104561, + 0.0361325666308403, + 0.0225394144654274, + 0.025728661566972733, + 0.017571084201335907, + 0.01239799614995718, + 0.01672084629535675, + 0.01088081393390894, + 0.00722739240154624, + 0.012319198809564114, + 0.01079113595187664, + 0.01278004702180624, + 0.01023498922586441, + 0.008077753707766533, + 0.015631135553121567, + 0.008529501035809517, + 0.004431628156453371, + 0.007383464835584164, + 0.011508396826684475, + 0.009517635218799114, + 0.008973419666290283, + 0.013359577395021915, + 0.016094474121928215, + 0.010110077448189259, + 0.00582470977678895, + 0.007967899553477764, + 0.011671722866594791, + 0.010479072108864784, + 0.007828441448509693, + 0.018986890092492104, + 0.009693476371467113, + 0.00412738835439086, + 0.003694899147376418, + 0.004614876117557287, + 0.004429349210113287, + 0.00612063379958272, + 0.009102784097194672, + 0.00752126332372427, + 0.01075279526412487, + 0.005895333364605904, + 0.004262118600308895, + 0.006544583477079868, + 0.00562552735209465, + 0.006669334135949612, + 0.009368157014250755, + 0.0059699914418160915, + 0.03932502493262291, + 0.014019091613590717, + 0.00782759115099907, + 0.006574699655175209, + 0.01922231912612915, + 0.013787665404379368, + 0.012321658432483673, + 0.029835745692253113, + 0.005759728141129017, + 0.004426958970725536, + 0.003337644739076495, + 0.004626686219125986, + 0.005043136887252331, + 0.004731170833110809, + 0.0030079465359449387, + 0.004694122355431318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.014201059006154537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328200101852417 + ], + "opposite_logprob_tokens": [ + -0.7578200101852417 + ], + "logprob_reward": -0.6328200101852417, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328200101852417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13682597875595093, + 0.11008980870246887, + 0.21523626148700714, + 0.1039491519331932, + 0.09619151055812836, + 0.0999155044555664, + 0.056817177683115005, + 0.07411301136016846, + 0.03818072751164436, + 0.06868094205856323 + ], + "credit_map": [ + 0.021059300750494003, + 0.017475398257374763, + 0.015810130164027214, + 0.016346776857972145, + 0.01974036917090416, + 0.021003741770982742, + 0.011377204209566116, + 0.014013050124049187, + 0.014270205982029438, + 0.013320627622306347, + 0.012365843169391155, + 0.019352853298187256, + 0.011757607571780682, + 0.01784873753786087, + 0.011752857826650143, + 0.00942107755690813, + 0.03492938354611397, + 0.025174260139465332, + 0.018174076452851295, + 0.026530610397458076, + 0.027494067326188087, + 0.03520110249519348, + 0.02037692815065384, + 0.027355827391147614, + 0.016721973195672035, + 0.010046472772955894, + 0.007340069394558668, + 0.019877685233950615, + 0.010992947965860367, + 0.012892222963273525, + 0.010197858326137066, + 0.015879927203059196, + 0.019370177760720253, + 0.009408225305378437, + 0.007948537357151508, + 0.01484630350023508, + 0.0070584299974143505, + 0.009347857907414436, + 0.01365649700164795, + 0.014555488713085651, + 0.0239515732973814, + 0.011253418400883675, + 0.005475331097841263, + 0.007580189034342766, + 0.01327600795775652, + 0.010770324617624283, + 0.005567599553614855, + 0.02204105444252491, + 0.01141936145722866, + 0.005512007977813482, + 0.004278422798961401, + 0.007287857588380575, + 0.004824602510780096, + 0.006222584750503302, + 0.011086679995059967, + 0.006185657810419798, + 0.016430500894784927, + 0.006246841978281736, + 0.005764796398580074, + 0.008374533616006374, + 0.008463880978524685, + 0.007214761804789305, + 0.009844956919550896, + 0.011772742494940758, + 0.00561201898381114, + 0.005061795935034752, + 0.003350598970428109, + 0.003453661687672138, + 0.0074028088711202145, + 0.0046630678698420525, + 0.004242720548063517, + 0.004394054878503084, + 0.013428338803350925, + 0.00688675232231617, + 0.005801551043987274, + 0.009929623454809189, + 0.00551917077973485, + 0.007370972074568272, + 0.011645435355603695, + 0.008099101483821869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.6328200101852417 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a lush, hand-drawn style characteristic of '90s Disney animation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a red-eyed tree frog with smooth green skin and bright orange feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the frog positioned almost completely flat on a large, dew-covered tropical leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the frog hiding behind a dense, emerald green stem with its eyes wide in a look of panic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight creating warm golden light and soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583942532539368 + ], + "opposite_logprob_tokens": [ + -3.8955838680267334 + ], + "logprob_reward": -0.020583942532539368, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583942532539368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28665491938591003, + 0.16779030859470367, + 0.10935160517692566, + 0.07321391999721527, + 0.07757387310266495, + 0.08610612899065018, + 0.03783959522843361, + 0.04593568295240402, + 0.04274182766675949, + 0.07279214262962341 + ], + "credit_map": [ + 0.04807343706488609, + 0.042279474437236786, + 0.0264425128698349, + 0.027355050668120384, + 0.04065489023923874, + 0.05528618395328522, + 0.017657915130257607, + 0.028905468061566353, + 0.02285093255341053, + 0.027905110269784927, + 0.015836408361792564, + 0.027307849377393723, + 0.02174178883433342, + 0.024262983351945877, + 0.013731822371482849, + 0.014153418131172657, + 0.012801934964954853, + 0.00983734242618084, + 0.014949115924537182, + 0.02035260573029518, + 0.012567956931889057, + 0.014976962469518185, + 0.013326258398592472, + 0.010539430193603039, + 0.010345889255404472, + 0.007765212096273899, + 0.007369273342192173, + 0.009177965112030506, + 0.008506665006279945, + 0.01041845977306366, + 0.012928960844874382, + 0.006701490376144648, + 0.01623755879700184, + 0.009149682708084583, + 0.005453458521515131, + 0.004619955085217953, + 0.01334493700414896, + 0.011794209480285645, + 0.00453695235773921, + 0.0124371238052845, + 0.01918926276266575, + 0.008651933632791042, + 0.004937083926051855, + 0.0050573041662573814, + 0.014031970873475075, + 0.0108191454783082, + 0.007717943750321865, + 0.01570148579776287, + 0.0061046164482831955, + 0.0038678289856761694, + 0.002979306038469076, + 0.004012413322925568, + 0.006519716698676348, + 0.007414898835122585, + 0.00390082155354321, + 0.00303999288007617, + 0.010831711813807487, + 0.004572026431560516, + 0.003154021454975009, + 0.00244582025334239, + 0.007239001337438822, + 0.005640288814902306, + 0.0030405500438064337, + 0.009012260474264622, + 0.010054337792098522, + 0.00321090966463089, + 0.002946017775684595, + 0.0044129579328000546, + 0.004056832287460566, + 0.004898960702121258, + 0.008975261822342873, + 0.004186546895653009, + 0.014279407449066639, + 0.007212514989078045, + 0.005216648802161217, + 0.005439658649265766, + 0.01006125658750534, + 0.009707758203148842, + 0.008727244101464748, + 0.012147655710577965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.020583942532539368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_0/index_tensor([2872, 2872])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_0/index_tensor([2872, 2872])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..658d156cb4eaf4d3721876c683cce51ad34199e9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_0/index_tensor([2872, 2872])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a scene of profound tranquility, embodying traditional Japanese aesthetics. A thick, glowing plume of steam rises slowly and gracefully from a natural hot spring, its perimeter lined with dark, smooth river rocks that shimmer under the dual light sources. A large, pale full moon begins its majestic ascent into the deep blue night sky, its cool, silvery light catching the edges of the billowing steam, making it appear ethereal and weightless. Nearby, a classic granite stone lantern (*tðr*) casts a warm, shimmering amber glow, its soft light dancing across the wet rocks, creating a welcoming pool. This warm light also softly illuminates the roots and gnarled outline of a leafless maple tree, its delicate branches sharply silhouetted against the moonlit sky. The entire work is a study in contrasts—cool moonlight against the warm lantern fire, dark stone against the white steam—evoking a serene, contemplative atmosphere, the only perceptible movement being the gentle swirls of steam and the soft flicker of the lantern flame.", + "seed": 21573, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_0/index_2872_seed_21573_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025756172300316394, + "vlm_logprob": -0.025756172300316394, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.025756172300316394, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044139355421066284 + ], + "opposite_logprob_tokens": [ + -3.1691393852233887 + ], + "logprob_reward": -0.044139355421066284, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.044139355421066284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36796075105667114, + 0.1602872908115387, + 0.07015790045261383, + 0.06729073822498322, + 0.05795445293188095, + 0.0731644481420517, + 0.03165590018033981, + 0.02661488577723503, + 0.08232618868350983, + 0.06258746236562729 + ], + "credit_map": [ + 0.03219308704137802, + 0.04058893769979477, + 0.05969758704304695, + 0.02911657840013504, + 0.05621325597167015, + 0.07086028903722763, + 0.025305448099970818, + 0.05398552119731903, + 0.010952259413897991, + 0.012523000128567219, + 0.01843043975532055, + 0.011670767329633236, + 0.02088707685470581, + 0.03624983876943588, + 0.02545241266489029, + 0.024121489375829697, + 0.00743909040465951, + 0.011901921592652798, + 0.0074912626296281815, + 0.005999553948640823, + 0.009608836844563484, + 0.010353988967835903, + 0.008403699845075607, + 0.008959541097283363, + 0.0058405110612511635, + 0.00675221486017108, + 0.011310294270515442, + 0.006387223955243826, + 0.011634753085672855, + 0.011461766436696053, + 0.004225692246109247, + 0.009678277187049389, + 0.005502315238118172, + 0.005858182441443205, + 0.011742240749299526, + 0.006380675360560417, + 0.008445748127996922, + 0.007620686199516058, + 0.0045966715551912785, + 0.0078079234808683395, + 0.006901118438690901, + 0.006235851906239986, + 0.015734340995550156, + 0.008857140317559242, + 0.009781739674508572, + 0.009287271648645401, + 0.006539218593388796, + 0.009827757254242897, + 0.0030540458392351866, + 0.002827004063874483, + 0.006658988073468208, + 0.0037560067139565945, + 0.003464257577434182, + 0.00341506814584136, + 0.003991218283772469, + 0.00448931148275733, + 0.0023343630600720644, + 0.0024989836383610964, + 0.0031678418163210154, + 0.0021213877480477095, + 0.0047973692417144775, + 0.004070404451340437, + 0.003525602864101529, + 0.004098930396139622, + 0.009471575729548931, + 0.0067718434147536755, + 0.009468707256019115, + 0.009622400626540184, + 0.010961964726448059, + 0.01756729744374752, + 0.007397819776087999, + 0.011064566671848297, + 0.0055123488418757915, + 0.0069303084164857864, + 0.007634390145540237, + 0.007170289754867554, + 0.010622932575643063, + 0.01074590627104044, + 0.006222445517778397, + 0.00774883758276701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.044139355421066284 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001892477273941 + ], + "opposite_logprob_tokens": [ + -2.695019006729126 + ], + "logprob_reward": -0.07001892477273941, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001892477273941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32023361325263977, + 0.15029345452785492, + 0.07300569117069244, + 0.09325889497995377, + 0.07110793888568878, + 0.071211077272892, + 0.03795865923166275, + 0.02665030024945736, + 0.08237933367490768, + 0.07390099763870239 + ], + "credit_map": [ + 0.026789261028170586, + 0.025776971131563187, + 0.05950111523270607, + 0.0258396677672863, + 0.048867881298065186, + 0.06446671485900879, + 0.01907942444086075, + 0.04991260915994644, + 0.009594750590622425, + 0.009385444223880768, + 0.01830117218196392, + 0.011339562013745308, + 0.0187277439981699, + 0.034212831407785416, + 0.025464897975325584, + 0.02326705865561962, + 0.006002484820783138, + 0.00962289422750473, + 0.0070571135729551315, + 0.005052885506302118, + 0.009342029690742493, + 0.012813806533813477, + 0.012675720266997814, + 0.010438758879899979, + 0.008024687878787518, + 0.008341658860445023, + 0.012930813245475292, + 0.00734149431809783, + 0.016991812735795975, + 0.020012222230434418, + 0.004848654381930828, + 0.01476755365729332, + 0.00575310830026865, + 0.005626253318041563, + 0.01499644760042429, + 0.005997008644044399, + 0.012902727350592613, + 0.011672678403556347, + 0.003727750154212117, + 0.010431970469653606, + 0.00616449723020196, + 0.004989991895854473, + 0.01301262341439724, + 0.006673125084489584, + 0.01143032405525446, + 0.013190142810344696, + 0.005447276867926121, + 0.010303097777068615, + 0.0036624479107558727, + 0.0026570854242891073, + 0.007388530299067497, + 0.004468374419957399, + 0.0037745984736829996, + 0.005062602460384369, + 0.005642473232001066, + 0.005302548874169588, + 0.002783079631626606, + 0.0028047726955264807, + 0.0025669040624052286, + 0.0018210718408226967, + 0.005190335679799318, + 0.004887253977358341, + 0.0024737720377743244, + 0.004123110789805651, + 0.010164021514356136, + 0.006907771807163954, + 0.010572146624326706, + 0.010830837301909924, + 0.009033480659127235, + 0.01563182845711708, + 0.008547660894691944, + 0.010691593401134014, + 0.006507433019578457, + 0.008058357052505016, + 0.009115944616496563, + 0.007927872240543365, + 0.012814583256840706, + 0.013984430581331253, + 0.005857790820300579, + 0.009634589776396751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07001892477273941 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013429442187771201 + ], + "opposite_logprob_tokens": [ + -6.6263427734375 + ], + "logprob_reward": -0.0013429442187771201, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013429442187771201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3392263352870941, + 0.20118413865566254, + 0.10044725239276886, + 0.05316084995865822, + 0.037510670721530914, + 0.06565520912408829, + 0.02983997017145157, + 0.039819590747356415, + 0.07780582457780838, + 0.055350229144096375 + ], + "credit_map": [ + 0.04706825315952301, + 0.053482040762901306, + 0.04952741786837578, + 0.03926512598991394, + 0.03259987011551857, + 0.03944871202111244, + 0.03612076863646507, + 0.04171412065625191, + 0.020846160128712654, + 0.029125358909368515, + 0.021927300840616226, + 0.01502755843102932, + 0.02632237784564495, + 0.0467321053147316, + 0.012072419747710228, + 0.02913082204759121, + 0.011453310027718544, + 0.019511910155415535, + 0.007906164973974228, + 0.006974202115088701, + 0.01250172033905983, + 0.01585714891552925, + 0.013449720107018948, + 0.01279306598007679, + 0.005439469125121832, + 0.007799938786774874, + 0.007958720438182354, + 0.00536942295730114, + 0.007222854997962713, + 0.00759920384734869, + 0.005166063085198402, + 0.006605166010558605, + 0.004705092869699001, + 0.008198941126465797, + 0.00337808346375823, + 0.003111884230747819, + 0.005083885509520769, + 0.006010422483086586, + 0.0027010098565369844, + 0.004321345593780279, + 0.004415570292621851, + 0.006466037593781948, + 0.00923904124647379, + 0.00605021370574832, + 0.009573806077241898, + 0.0076412237249314785, + 0.011584264226257801, + 0.010685041546821594, + 0.003457512939348817, + 0.0034957730676978827, + 0.0042755743488669395, + 0.003837812226265669, + 0.0026493463665246964, + 0.004833124577999115, + 0.003496376797556877, + 0.003794447286054492, + 0.004109405912458897, + 0.00468683335930109, + 0.004153016023337841, + 0.004480751231312752, + 0.006295234430581331, + 0.0055669741705060005, + 0.004990467801690102, + 0.005536903161555529, + 0.008710642345249653, + 0.006791811902076006, + 0.011458226479589939, + 0.00777908880263567, + 0.008999580517411232, + 0.01241198368370533, + 0.010686827823519707, + 0.010967649519443512, + 0.0073991334065794945, + 0.009806240908801556, + 0.007622475735843182, + 0.005739915184676647, + 0.00576084153726697, + 0.006736009381711483, + 0.006405920721590519, + 0.005879685748368502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0013429442187771201 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021969962399452925 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021969962399452925, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021969962399452925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4303264915943146, + 0.18665318191051483, + 0.14627335965633392, + 0.027654947713017464, + 0.024960719048976898, + 0.04590914770960808, + 0.02827998250722885, + 0.02786472626030445, + 0.05262172594666481, + 0.02945566549897194 + ], + "credit_map": [ + 0.03339513763785362, + 0.051633212715387344, + 0.055789314210414886, + 0.03713975101709366, + 0.04490036889910698, + 0.0794227123260498, + 0.07132510840892792, + 0.056720927357673645, + 0.014077449217438698, + 0.014207831583917141, + 0.02466670423746109, + 0.01643143594264984, + 0.021494589745998383, + 0.04341818392276764, + 0.02480381168425083, + 0.027553189545869827, + 0.00870547816157341, + 0.014929039403796196, + 0.013535680249333382, + 0.012422023341059685, + 0.014544829726219177, + 0.0256680678576231, + 0.03424954414367676, + 0.022218700498342514, + 0.0024201027117669582, + 0.0027495892718434334, + 0.004718378186225891, + 0.0033646163064986467, + 0.003104999428614974, + 0.004474605433642864, + 0.003171872813254595, + 0.003650784259662032, + 0.002146020531654358, + 0.0028171143494546413, + 0.004272678401321173, + 0.0031568545382469893, + 0.0031512193381786346, + 0.003661505179479718, + 0.002646715845912695, + 0.003108612261712551, + 0.0030049323104321957, + 0.005308981519192457, + 0.007385812699794769, + 0.00496421754360199, + 0.006100633181631565, + 0.004285627044737339, + 0.008040010929107666, + 0.006818937603384256, + 0.0027511154767125845, + 0.003378036431968212, + 0.002848915522918105, + 0.0020643165335059166, + 0.005209536291658878, + 0.005446046590805054, + 0.002913060365244746, + 0.0036689587868750095, + 0.003377731190994382, + 0.003058808157220483, + 0.004119187127798796, + 0.0031765878666192293, + 0.0029240604490041733, + 0.002944602631032467, + 0.0037659467197954655, + 0.004497803747653961, + 0.00597643107175827, + 0.004065629094839096, + 0.007229464128613472, + 0.0043716514483094215, + 0.0075227865017950535, + 0.010260997340083122, + 0.005848440807312727, + 0.007346328813582659, + 0.003229688387364149, + 0.0035063421819359064, + 0.003384316572919488, + 0.003289279527962208, + 0.004208661615848541, + 0.005265176296234131, + 0.0031983775552362204, + 0.003373826155439019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0021969962399452925 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108264084905386 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.01108264084905386, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108264084905386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2626460790634155, + 0.08488990366458893, + 0.12269304692745209, + 0.07508384436368942, + 0.08275166153907776, + 0.07151869684457779, + 0.07173867523670197, + 0.05354295298457146, + 0.10363460332155228, + 0.07150045782327652 + ], + "credit_map": [ + 0.027253244072198868, + 0.027878599241375923, + 0.05709005519747734, + 0.03774455189704895, + 0.0190576184540987, + 0.03084813430905342, + 0.02801617421209812, + 0.034757744520902634, + 0.008091838099062443, + 0.010326696559786797, + 0.009914922527968884, + 0.013717243447899818, + 0.009935924783349037, + 0.010908755473792553, + 0.01140392106026411, + 0.010590611957013607, + 0.009764383547008038, + 0.015576270408928394, + 0.011608385480940342, + 0.008733374066650867, + 0.018285058438777924, + 0.029997678473591805, + 0.009761746972799301, + 0.01896616443991661, + 0.006524510681629181, + 0.010465365834534168, + 0.0067180306650698185, + 0.0065744235180318356, + 0.009283791296184063, + 0.00838372390717268, + 0.01644831895828247, + 0.010685688816010952, + 0.007573409471660852, + 0.011240239255130291, + 0.008278134278953075, + 0.011003907769918442, + 0.01188002247363329, + 0.01085385587066412, + 0.011376460082828999, + 0.010545639321208, + 0.006314076017588377, + 0.00785856507718563, + 0.007593069225549698, + 0.00660408241674304, + 0.009796448051929474, + 0.013292462565004826, + 0.008855714462697506, + 0.01120428740978241, + 0.008255807682871819, + 0.010522457771003246, + 0.009309782646596432, + 0.005870826542377472, + 0.01097051240503788, + 0.009829620830714703, + 0.007637325208634138, + 0.009342349134385586, + 0.005624727346003056, + 0.005440101958811283, + 0.008477677591145039, + 0.0054939910769462585, + 0.006855637766420841, + 0.009425250813364983, + 0.004745540674775839, + 0.0074800290167331696, + 0.009250488132238388, + 0.01327070128172636, + 0.014137009158730507, + 0.010583525523543358, + 0.014287690632045269, + 0.01213323324918747, + 0.015152910724282265, + 0.014819049276411533, + 0.007282710168510675, + 0.013544042594730854, + 0.009678122587502003, + 0.0071331895887851715, + 0.008176567032933235, + 0.009281204082071781, + 0.007793638855218887, + 0.008610996417701244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.01108264084905386 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a scene of profound tranquility, embodying traditional Japanese aesthetics. A thick, glowing plume of steam rises slowly and gracefully from a natural hot spring, its perimeter lined with dark, smooth river rocks that shimmer under the dual light sources. A large, pale full moon begins its majestic ascent into the deep blue night sky, its cool, silvery light catching the edges of the billowing steam, making it appear ethereal and weightless. Nearby, a classic granite stone lantern (*tðr*) casts a warm, shimmering amber glow, its soft light dancing across the wet rocks, creating a welcoming pool. This warm light also softly illuminates the roots and gnarled outline of a leafless maple tree, its delicate branches sharply silhouetted against the moonlit sky. The entire work is a study in contrasts—cool moonlight against the warm lantern fire, dark stone against the white steam—evoking a serene, contemplative atmosphere, the only perceptible movement being the gentle swirls of steam and the soft flicker of the lantern flame.", + "seed": 21574, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_0/index_2872_seed_21574_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.009611792128998787, + "vlm_logprob": -0.009611792128998787, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.009611792128998787, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005571551155298948 + ], + "opposite_logprob_tokens": [ + -5.380571365356445 + ], + "logprob_reward": -0.005571551155298948, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.005571551155298948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2705693542957306, + 0.06750429421663284, + 0.09917128086090088, + 0.10338883101940155, + 0.032768018543720245, + 0.02584701031446457, + 0.02574363723397255, + 0.031224701553583145, + 0.057122547179460526, + 0.28666025400161743 + ], + "credit_map": [ + 0.025374555960297585, + 0.040998246520757675, + 0.0354745127260685, + 0.034575287252664566, + 0.0407220833003521, + 0.03676948323845863, + 0.03428035229444504, + 0.022374846041202545, + 0.005650877952575684, + 0.01165600772947073, + 0.005319397896528244, + 0.005269971676170826, + 0.010497676208615303, + 0.01454621646553278, + 0.00481779221445322, + 0.009746357798576355, + 0.011531558819115162, + 0.017261900007724762, + 0.008380108512938023, + 0.007207080721855164, + 0.015814276412129402, + 0.01974124275147915, + 0.008557415567338467, + 0.010677694343030453, + 0.01220633089542389, + 0.025773806497454643, + 0.008163193240761757, + 0.007065273821353912, + 0.01358780451118946, + 0.018253566697239876, + 0.005844795610755682, + 0.012494061142206192, + 0.0034704203717410564, + 0.004715878516435623, + 0.0037155430763959885, + 0.003011044580489397, + 0.005358410533517599, + 0.005858602933585644, + 0.00228809448890388, + 0.00435002613812685, + 0.002035661833360791, + 0.0020608382765203714, + 0.0036754359025508165, + 0.003411895828321576, + 0.0032266885973513126, + 0.0028413645923137665, + 0.0046060471795499325, + 0.00398907670751214, + 0.0031321330461651087, + 0.0021471199579536915, + 0.003364170901477337, + 0.0027835669461637735, + 0.004080094397068024, + 0.003598008770495653, + 0.0027675889432430267, + 0.0038709542714059353, + 0.0027658476028591394, + 0.0032702693715691566, + 0.003276229603216052, + 0.0029174808878451586, + 0.004547069780528545, + 0.0061472817324101925, + 0.0033483835868537426, + 0.004952138289809227, + 0.004542363807559013, + 0.003934200387448072, + 0.0058067976497113705, + 0.0068741231225430965, + 0.009232853539288044, + 0.008769921958446503, + 0.009408409707248211, + 0.008553877472877502, + 0.02206408977508545, + 0.03595306724309921, + 0.021220536902546883, + 0.0371532179415226, + 0.04171346500515938, + 0.023580161854624748, + 0.062379151582717896, + 0.042596567422151566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.005571551155298948 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285161703824997 + ], + "opposite_logprob_tokens": [ + -3.773285150527954 + ], + "logprob_reward": -0.023285161703824997, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285161703824997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2981607913970947, + 0.07385864853858948, + 0.08623660355806351, + 0.05599731579422951, + 0.056454747915267944, + 0.041462406516075134, + 0.031742896884679794, + 0.045447248965501785, + 0.038093145936727524, + 0.2725462317466736 + ], + "credit_map": [ + 0.019214091822504997, + 0.029528189450502396, + 0.03268572688102722, + 0.04023753106594086, + 0.04524831101298332, + 0.04053113982081413, + 0.051204897463321686, + 0.03951089084148407, + 0.006013704929500818, + 0.011005710810422897, + 0.006997672840952873, + 0.008435933850705624, + 0.012342801317572594, + 0.009370972402393818, + 0.010053873993456364, + 0.00963797140866518, + 0.006299776025116444, + 0.01186918281018734, + 0.007585263345390558, + 0.009254037402570248, + 0.016263974830508232, + 0.014468417502939701, + 0.009169482626020908, + 0.01132646482437849, + 0.006382504943758249, + 0.008357464335858822, + 0.0032436675392091274, + 0.00437556765973568, + 0.009680519811809063, + 0.011357193812727928, + 0.005053365137428045, + 0.00754702789708972, + 0.005924866069108248, + 0.006691563408821821, + 0.005597820971161127, + 0.005482422187924385, + 0.00831434316933155, + 0.010485521517693996, + 0.005717682186514139, + 0.00824052095413208, + 0.003886203048750758, + 0.00539670092985034, + 0.005084119737148285, + 0.004802908282727003, + 0.005461159627884626, + 0.005069849546998739, + 0.006128290668129921, + 0.005633174441754818, + 0.003271020483225584, + 0.0034886409994214773, + 0.002890486503019929, + 0.0028816568665206432, + 0.0067662508226931095, + 0.004861572291702032, + 0.002449828665703535, + 0.005133436527103186, + 0.00374225084669888, + 0.008156896568834782, + 0.005256368778645992, + 0.008104170672595501, + 0.005096616689115763, + 0.005206374451518059, + 0.0042651002295315266, + 0.00561946677044034, + 0.0036726053804159164, + 0.003312057117000222, + 0.003594981273636222, + 0.00467253290116787, + 0.004658542573451996, + 0.005224017892032862, + 0.007532263174653053, + 0.00542614609003067, + 0.02002807706594467, + 0.032718777656555176, + 0.019964290782809258, + 0.03644280880689621, + 0.038080327212810516, + 0.02235708385705948, + 0.06182565540075302, + 0.04112919047474861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.023285161703824997 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022038952447474003 + ], + "opposite_logprob_tokens": [ + -6.127203941345215 + ], + "logprob_reward": -0.0022038952447474003, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022038952447474003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11570949107408524, + 0.14194315671920776, + 0.09064813703298569, + 0.20783032476902008, + 0.07876861095428467, + 0.040663715451955795, + 0.05211162194609642, + 0.025304635986685753, + 0.03306347504258156, + 0.21395686268806458 + ], + "credit_map": [ + 0.009874640963971615, + 0.014659902080893517, + 0.01173167210072279, + 0.010777454823255539, + 0.031567130237817764, + 0.016129270195961, + 0.010717472061514854, + 0.010251948609948158, + 0.014916460029780865, + 0.039320871233940125, + 0.011773288249969482, + 0.006947461981326342, + 0.02816738188266754, + 0.02292669005692005, + 0.006831231992691755, + 0.011059770360589027, + 0.0100995022803545, + 0.018416445702314377, + 0.007397839799523354, + 0.005635169800370932, + 0.018169675022363663, + 0.014143493957817554, + 0.0071773836389184, + 0.00960863009095192, + 0.026101207360625267, + 0.06460335850715637, + 0.0172694381326437, + 0.01746351458132267, + 0.025037439540028572, + 0.025961212813854218, + 0.01130182296037674, + 0.020092325285077095, + 0.008953100070357323, + 0.019827384501695633, + 0.006718727760016918, + 0.005308155901730061, + 0.013948680832982063, + 0.011004919186234474, + 0.0041039567440748215, + 0.00890368688851595, + 0.005717654246836901, + 0.008475026115775108, + 0.004597814287990332, + 0.0024644029326736927, + 0.007129458244889975, + 0.00474525336176157, + 0.0024389801546931267, + 0.0050951275043189526, + 0.005116398446261883, + 0.008063345216214657, + 0.004914978519082069, + 0.006241101771593094, + 0.006959106773138046, + 0.00792978797107935, + 0.005588156636804342, + 0.007298745214939117, + 0.0026769505348056555, + 0.003495848271995783, + 0.003498176811262965, + 0.0036954754032194614, + 0.0034623793326318264, + 0.003124580252915621, + 0.0022028384264558554, + 0.0031483876518905163, + 0.0033786268904805183, + 0.0025755867827683687, + 0.003542540827766061, + 0.004946415312588215, + 0.004410552326589823, + 0.004061732906848192, + 0.0051887803710997105, + 0.0049592372961342335, + 0.015965064987540245, + 0.02168918214738369, + 0.014392384327948093, + 0.02995484136044979, + 0.030997227877378464, + 0.01749117113649845, + 0.04939122498035431, + 0.03407576307654381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.0022038952447474003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009214208112098277 + ], + "opposite_logprob_tokens": [ + -7.000921249389648 + ], + "logprob_reward": -0.0009214208112098277, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009214208112098277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061908137053251266, + 0.05164003372192383, + 0.06058908626437187, + 0.039212994277477264, + 0.057581476867198944, + 0.03897295892238617, + 0.02992582693696022, + 0.04814165085554123, + 0.1657486855983734, + 0.4462791383266449 + ], + "credit_map": [ + 0.007221336476504803, + 0.007327148225158453, + 0.00744028203189373, + 0.008977358229458332, + 0.010862360708415508, + 0.007149284705519676, + 0.007805184926837683, + 0.005125180818140507, + 0.0048639411106705666, + 0.006115710828453302, + 0.004021105356514454, + 0.004250552970916033, + 0.00816279835999012, + 0.011296895332634449, + 0.004553689621388912, + 0.008375340141355991, + 0.007626429200172424, + 0.004963588900864124, + 0.0030670317355543375, + 0.002987524261698127, + 0.01347150094807148, + 0.016602635383605957, + 0.0043510873802006245, + 0.007519292179495096, + 0.0032190666534006596, + 0.004291334189474583, + 0.003330857027322054, + 0.004305734299123287, + 0.006158194504678249, + 0.007436997257173061, + 0.005028276238590479, + 0.0054425341077148914, + 0.004344515036791563, + 0.004150536842644215, + 0.005196402780711651, + 0.0048240432515740395, + 0.009405359625816345, + 0.011565993539988995, + 0.009239074774086475, + 0.008855551481246948, + 0.0035026976838707924, + 0.004938877187669277, + 0.0033739348873496056, + 0.002719081938266754, + 0.006686611101031303, + 0.007030236069113016, + 0.005012122448533773, + 0.005709400400519371, + 0.00288296933285892, + 0.0025786771439015865, + 0.002701800549402833, + 0.0034047330264002085, + 0.005672710947692394, + 0.005200404208153486, + 0.0032687014900147915, + 0.004215829540044069, + 0.004428582265973091, + 0.006400248501449823, + 0.005828466732054949, + 0.005950799211859703, + 0.006636367179453373, + 0.006154660601168871, + 0.005725752096623182, + 0.007016776129603386, + 0.014960180036723614, + 0.016705641523003578, + 0.014494859613478184, + 0.020703554153442383, + 0.023065796121954918, + 0.017534606158733368, + 0.03440637141466141, + 0.023877684026956558, + 0.03232528641819954, + 0.04842953011393547, + 0.03290944918990135, + 0.06194030120968819, + 0.06201493367552757, + 0.03455622121691704, + 0.10475154221057892, + 0.06935185194015503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0009214208112098277 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076931729912758 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.016076931729912758, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076931729912758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11523096263408661, + 0.07005131244659424, + 0.0666019469499588, + 0.05439621955156326, + 0.0681193545460701, + 0.03508920222520828, + 0.03922824189066887, + 0.05657810717821121, + 0.11648417264223099, + 0.37822046875953674 + ], + "credit_map": [ + 0.011610451154410839, + 0.018028125166893005, + 0.011945796199142933, + 0.010480481199920177, + 0.027563530951738358, + 0.016273440793156624, + 0.008982304483652115, + 0.010346838273108006, + 0.007355895359069109, + 0.005684466566890478, + 0.00502355070784688, + 0.00415548263117671, + 0.00921886507421732, + 0.019098786637187004, + 0.006091138813644648, + 0.013423124328255653, + 0.005758850369602442, + 0.01274500135332346, + 0.004738117102533579, + 0.0037941548507660627, + 0.014033862389624119, + 0.013164008967578411, + 0.004847944714128971, + 0.007520005572587252, + 0.004181828815490007, + 0.008457657881081104, + 0.003528477856889367, + 0.0035682055167853832, + 0.009200740605592728, + 0.012431193143129349, + 0.00518786208704114, + 0.007840254344046116, + 0.006333289202302694, + 0.013360857032239437, + 0.006078946869820356, + 0.006075642537325621, + 0.014018089510500431, + 0.009336933493614197, + 0.004456155002117157, + 0.008459438569843769, + 0.0028742672875523567, + 0.0027807638980448246, + 0.0036825237330049276, + 0.0027896214742213488, + 0.005845389328896999, + 0.00757726700976491, + 0.004500098060816526, + 0.005039272829890251, + 0.004120387136936188, + 0.005013075191527605, + 0.003751506796106696, + 0.0031847951468080282, + 0.008132584393024445, + 0.005872732028365135, + 0.0036485951859503984, + 0.005504565313458443, + 0.003761640517041087, + 0.007379072718322277, + 0.0040718382224440575, + 0.0060095214284956455, + 0.010883471928536892, + 0.009692518040537834, + 0.005880680866539478, + 0.008899365551769733, + 0.010643589310348034, + 0.017660239711403847, + 0.012220307253301144, + 0.01186071615666151, + 0.013950005173683167, + 0.012098738923668861, + 0.02227061614394188, + 0.015779957175254822, + 0.02835523895919323, + 0.04256420210003853, + 0.028324181213974953, + 0.05125981569290161, + 0.0528876893222332, + 0.030651237815618515, + 0.08655180037021637, + 0.057626307010650635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.016076931729912758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a scene of profound tranquility, embodying traditional Japanese aesthetics. A thick, glowing plume of steam rises slowly and gracefully from a natural hot spring, its perimeter lined with dark, smooth river rocks that shimmer under the dual light sources. A large, pale full moon begins its majestic ascent into the deep blue night sky, its cool, silvery light catching the edges of the billowing steam, making it appear ethereal and weightless. Nearby, a classic granite stone lantern (*tðr*) casts a warm, shimmering amber glow, its soft light dancing across the wet rocks, creating a welcoming pool. This warm light also softly illuminates the roots and gnarled outline of a leafless maple tree, its delicate branches sharply silhouetted against the moonlit sky. The entire work is a study in contrasts—cool moonlight against the warm lantern fire, dark stone against the white steam—evoking a serene, contemplative atmosphere, the only perceptible movement being the gentle swirls of steam and the soft flicker of the lantern flame.", + "seed": 21575, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_0/index_2872_seed_21575_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014507451467216015, + "vlm_logprob": -0.014507451467216015, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.014507451467216015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010715564712882042 + ], + "opposite_logprob_tokens": [ + -4.635715484619141 + ], + "logprob_reward": -0.010715564712882042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010715564712882042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25929373502731323, + 0.06930355727672577, + 0.059067100286483765, + 0.05699275806546211, + 0.06247507408261299, + 0.10138920694589615, + 0.06399405002593994, + 0.07597941160202026, + 0.14700452983379364, + 0.10450058430433273 + ], + "credit_map": [ + 0.026033636182546616, + 0.017768071964383125, + 0.027852162718772888, + 0.026333987712860107, + 0.04682758077979088, + 0.03924332931637764, + 0.03397022560238838, + 0.041264720261096954, + 0.007437600288540125, + 0.00505484314635396, + 0.00958953332155943, + 0.007558487355709076, + 0.010286710225045681, + 0.0103612020611763, + 0.00788059737533331, + 0.011134574189782143, + 0.005712170619517565, + 0.002981786848977208, + 0.007051183842122555, + 0.005839490797370672, + 0.007153044920414686, + 0.0102959293872118, + 0.009532484225928783, + 0.010501008480787277, + 0.005207522306591272, + 0.003683480666950345, + 0.005105337128043175, + 0.0039241430349648, + 0.0056854961439967155, + 0.01274314895272255, + 0.010816231369972229, + 0.009827394038438797, + 0.0072196777909994125, + 0.005040677729994059, + 0.007782050874084234, + 0.0037914286367595196, + 0.007667969446629286, + 0.011255991645157337, + 0.01012029405683279, + 0.009596983902156353, + 0.00818546675145626, + 0.012662970460951328, + 0.009599873796105385, + 0.007746115792542696, + 0.012360084801912308, + 0.016218265518546104, + 0.019447676837444305, + 0.015168757177889347, + 0.009680879302322865, + 0.01382501795887947, + 0.005278256721794605, + 0.005739854648709297, + 0.006488020531833172, + 0.011037152260541916, + 0.003642084775492549, + 0.008302779868245125, + 0.010204863734543324, + 0.01446011383086443, + 0.006927849724888802, + 0.005688828881829977, + 0.009773580357432365, + 0.01219108048826456, + 0.007531323470175266, + 0.00920176412910223, + 0.01656813733279705, + 0.01881352812051773, + 0.011990436352789402, + 0.011248021386563778, + 0.028503738343715668, + 0.027259431779384613, + 0.014890175312757492, + 0.017731061205267906, + 0.009628905914723873, + 0.011485784314572811, + 0.010825286619365215, + 0.009821229614317417, + 0.027240145951509476, + 0.013568038120865822, + 0.009672543033957481, + 0.012258641421794891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.010715564712882042 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305346682667732 + ], + "opposite_logprob_tokens": [ + -3.168053388595581 + ], + "logprob_reward": -0.04305346682667732, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305346682667732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17490461468696594, + 0.07727718353271484, + 0.08106361329555511, + 0.05034433305263519, + 0.06662196666002274, + 0.11777646094560623, + 0.07528973370790482, + 0.1644410490989685, + 0.13162650167942047, + 0.06065450236201286 + ], + "credit_map": [ + 0.0172880832105875, + 0.0096604498103261, + 0.02378401719033718, + 0.018240563571453094, + 0.03200133889913559, + 0.02816559560596943, + 0.017610574141144753, + 0.028153996914625168, + 0.007083657663315535, + 0.004641041625291109, + 0.012925727292895317, + 0.008452830836176872, + 0.011188764125108719, + 0.011075343005359173, + 0.009386901743710041, + 0.012522917240858078, + 0.006381835788488388, + 0.003487908048555255, + 0.011371297761797905, + 0.004030013456940651, + 0.01533533912152052, + 0.017316700890660286, + 0.008021879009902477, + 0.01511863898485899, + 0.0044643450528383255, + 0.0025822282768785954, + 0.0050216978415846825, + 0.0031882114708423615, + 0.005733550060540438, + 0.010209533385932446, + 0.00969469640403986, + 0.009450065903365612, + 0.00509042851626873, + 0.005423674825578928, + 0.006962195970118046, + 0.003772948868572712, + 0.010306842625141144, + 0.012475624680519104, + 0.011825622990727425, + 0.010764630511403084, + 0.01094245444983244, + 0.008955907076597214, + 0.01323680579662323, + 0.009906898252665997, + 0.013866985216736794, + 0.01850847899913788, + 0.02436200901865959, + 0.017996923997998238, + 0.008437935262918472, + 0.01506041083484888, + 0.005742386449128389, + 0.005842478480190039, + 0.011811369098722935, + 0.01282386016100645, + 0.007429241668432951, + 0.008142050355672836, + 0.013823588378727436, + 0.01648220606148243, + 0.01367042027413845, + 0.011499292217195034, + 0.01628699339926243, + 0.050890691578388214, + 0.01391930878162384, + 0.02786855213344097, + 0.01349260937422514, + 0.008539145812392235, + 0.011445940472185612, + 0.009269166737794876, + 0.023733410984277725, + 0.03514464199542999, + 0.011426969431340694, + 0.018574612215161324, + 0.005783947184681892, + 0.005696070846170187, + 0.006461132317781448, + 0.005390671547502279, + 0.011300751939415932, + 0.010882659815251827, + 0.007393650710582733, + 0.0077456142753362656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.04305346682667732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015165030490607023 + ], + "opposite_logprob_tokens": [ + -6.501516342163086 + ], + "logprob_reward": -0.0015165030490607023, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015165030490607023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17561021447181702, + 0.04786877706646919, + 0.03673959895968437, + 0.03046342171728611, + 0.051977865397930145, + 0.11335498839616776, + 0.07341443747282028, + 0.1544792205095291, + 0.23052962124347687, + 0.0855618342757225 + ], + "credit_map": [ + 0.023698920384049416, + 0.012147802859544754, + 0.01899462752044201, + 0.014354667626321316, + 0.02827187068760395, + 0.0364190936088562, + 0.01495272945612669, + 0.02677052468061447, + 0.006055122707039118, + 0.003749377792701125, + 0.005818567704409361, + 0.003967511933296919, + 0.008335890248417854, + 0.007265669293701649, + 0.004442625213414431, + 0.008234012871980667, + 0.003619591472670436, + 0.002741072792559862, + 0.006877279840409756, + 0.002992139896377921, + 0.004622216802090406, + 0.005111054517328739, + 0.004795321263372898, + 0.005980923771858215, + 0.003585974918678403, + 0.00411834055557847, + 0.0035254356916993856, + 0.002884885296225548, + 0.0036346628330647945, + 0.004234577529132366, + 0.0036823630798608065, + 0.0047971829771995544, + 0.005773319397121668, + 0.0059499843046069145, + 0.006821951828896999, + 0.004712262190878391, + 0.009040690027177334, + 0.006120885256677866, + 0.00587105005979538, + 0.007687726058065891, + 0.01439477689564228, + 0.03091517835855484, + 0.009966941550374031, + 0.009046413004398346, + 0.013774576596915722, + 0.012083658017218113, + 0.011172467842698097, + 0.01200098730623722, + 0.010011686943471432, + 0.012792717665433884, + 0.005705118644982576, + 0.006360700819641352, + 0.012999499216675758, + 0.008782806806266308, + 0.00833783857524395, + 0.00842407438904047, + 0.025062069296836853, + 0.02859566919505596, + 0.015598731115460396, + 0.01169049832969904, + 0.025724897161126137, + 0.02075308933854103, + 0.00907551683485508, + 0.017978740856051445, + 0.030426248908042908, + 0.054388441145420074, + 0.018943827599287033, + 0.019964655861258507, + 0.03275199607014656, + 0.028107477352023125, + 0.019237641245126724, + 0.026709357276558876, + 0.010920541360974312, + 0.019453860819339752, + 0.006695597432553768, + 0.006320748012512922, + 0.013571251183748245, + 0.012013228610157967, + 0.006551350001245737, + 0.010035250335931778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0015165030490607023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011799524072557688 + ], + "opposite_logprob_tokens": [ + -6.751180171966553 + ], + "logprob_reward": -0.0011799524072557688, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011799524072557688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2838205397129059, + 0.07881271839141846, + 0.11336348205804825, + 0.06902454793453217, + 0.15824642777442932, + 0.09286865592002869, + 0.04052606225013733, + 0.030685121193528175, + 0.05861347168684006, + 0.07403893768787384 + ], + "credit_map": [ + 0.04366699606180191, + 0.017587708309292793, + 0.03491590917110443, + 0.024991456419229507, + 0.047754112631082535, + 0.04129979386925697, + 0.03034953586757183, + 0.0432550422847271, + 0.006456029135733843, + 0.005201912019401789, + 0.010218368843197823, + 0.007046010810881853, + 0.013344572857022285, + 0.013254004530608654, + 0.011709998361766338, + 0.011581824161112309, + 0.00846010446548462, + 0.004928744398057461, + 0.015568108297884464, + 0.009819046594202518, + 0.01671803742647171, + 0.018806803971529007, + 0.019185779616236687, + 0.019876856356859207, + 0.004272555932402611, + 0.0029001065995544195, + 0.009098480455577374, + 0.004834131337702274, + 0.009836230427026749, + 0.012509116902947426, + 0.013229714706540108, + 0.012344210408627987, + 0.011336698196828365, + 0.008897463791072369, + 0.013254476711153984, + 0.011446020565927029, + 0.02267911285161972, + 0.029249204322695732, + 0.03494330495595932, + 0.026440145447850227, + 0.008139904588460922, + 0.012852640822529793, + 0.01078865211457014, + 0.006063827313482761, + 0.013496872037649155, + 0.01589360646903515, + 0.013810454867780209, + 0.011822701431810856, + 0.004940051585435867, + 0.007224740460515022, + 0.003101175418123603, + 0.002884231274947524, + 0.006238377187401056, + 0.00765956612303853, + 0.0029587987810373306, + 0.005519118160009384, + 0.00413363566622138, + 0.003637289861217141, + 0.002489703707396984, + 0.004154484253376722, + 0.004241149872541428, + 0.0052331886254251, + 0.003230021335184574, + 0.00356564880348742, + 0.0070989057421684265, + 0.011645684018731117, + 0.00425098417326808, + 0.0057623498141765594, + 0.006143575068563223, + 0.012030431069433689, + 0.004181052558124065, + 0.007500488311052322, + 0.00795914325863123, + 0.011034521274268627, + 0.0055091967806220055, + 0.007142138667404652, + 0.014401691034436226, + 0.012596691027283669, + 0.004663776606321335, + 0.0107317715883255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.0011799524072557688 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607177034020424 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.01607177034020424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607177034020424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10938046127557755, + 0.08105924725532532, + 0.09593190252780914, + 0.0519416481256485, + 0.10303434729576111, + 0.11565820872783661, + 0.09821811318397522, + 0.06362324953079224, + 0.1336214393377304, + 0.1475314348936081 + ], + "credit_map": [ + 0.010657701641321182, + 0.007140145171433687, + 0.010737497359514236, + 0.008937055245041847, + 0.021843766793608665, + 0.018197020515799522, + 0.013260086067020893, + 0.018607178702950478, + 0.00933575164526701, + 0.005473204422742128, + 0.007232155650854111, + 0.008324408903717995, + 0.01279786229133606, + 0.015494965016841888, + 0.009063697420060635, + 0.013337195850908756, + 0.010151886381208897, + 0.004985491279512644, + 0.022250104695558548, + 0.011636952869594097, + 0.0059707812033593655, + 0.013934170827269554, + 0.009289151057600975, + 0.017713360488414764, + 0.003963956609368324, + 0.0031506104860454798, + 0.005953111685812473, + 0.004128291737288237, + 0.006989908870309591, + 0.0076705217361450195, + 0.01145921926945448, + 0.008626025170087814, + 0.008901757188141346, + 0.008976446464657784, + 0.012450759299099445, + 0.009456954896450043, + 0.013288254849612713, + 0.014941738918423653, + 0.01867666095495224, + 0.01634177379310131, + 0.012792330235242844, + 0.026553506031632423, + 0.011301595717668533, + 0.00794632826000452, + 0.013606115244328976, + 0.015461019240319729, + 0.013078424148261547, + 0.014918888919055462, + 0.01121615432202816, + 0.016398977488279343, + 0.007812161464244127, + 0.009909482672810555, + 0.014560490846633911, + 0.016175342723727226, + 0.009825860150158405, + 0.012319645844399929, + 0.009990077465772629, + 0.007830382324755192, + 0.004964231513440609, + 0.005403812043368816, + 0.009580946527421474, + 0.012065505608916283, + 0.005601849406957626, + 0.008186442777514458, + 0.019205205142498016, + 0.03192509338259697, + 0.0096068624407053, + 0.009827831760048866, + 0.016818003728985786, + 0.022972166538238525, + 0.007846930995583534, + 0.015419341623783112, + 0.013355162926018238, + 0.015339824371039867, + 0.013938277959823608, + 0.016534864902496338, + 0.026466792449355125, + 0.02259620651602745, + 0.020390236750245094, + 0.01891007460653782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.01607177034020424 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a scene of profound tranquility, embodying traditional Japanese aesthetics. A thick, glowing plume of steam rises slowly and gracefully from a natural hot spring, its perimeter lined with dark, smooth river rocks that shimmer under the dual light sources. A large, pale full moon begins its majestic ascent into the deep blue night sky, its cool, silvery light catching the edges of the billowing steam, making it appear ethereal and weightless. Nearby, a classic granite stone lantern (*tðr*) casts a warm, shimmering amber glow, its soft light dancing across the wet rocks, creating a welcoming pool. This warm light also softly illuminates the roots and gnarled outline of a leafless maple tree, its delicate branches sharply silhouetted against the moonlit sky. The entire work is a study in contrasts—cool moonlight against the warm lantern fire, dark stone against the white steam—evoking a serene, contemplative atmosphere, the only perceptible movement being the gentle swirls of steam and the soft flicker of the lantern flame.", + "seed": 21576, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_0/index_2872_seed_21576_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2372367882169783, + "vlm_logprob": -0.2372367882169783, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.2372367882169783, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.899097740650177 + ], + "opposite_logprob_tokens": [ + -0.524097740650177 + ], + "logprob_reward": -0.899097740650177, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.899097740650177, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2543394863605499, + 0.3994232714176178, + 0.024415655061602592, + 0.038122523576021194, + 0.08017858117818832, + 0.06781069189310074, + 0.0172248724848032, + 0.0519312247633934, + 0.03748813644051552, + 0.029065560549497604 + ], + "credit_map": [ + 0.025120574980974197, + 0.03151706978678703, + 0.040294647216796875, + 0.0337713398039341, + 0.02717115730047226, + 0.033936433494091034, + 0.025831596925854683, + 0.03669667989015579, + 0.05087277293205261, + 0.03663574531674385, + 0.07701218873262405, + 0.0348849855363369, + 0.04989543557167053, + 0.04528621956706047, + 0.04427509382367134, + 0.06056080013513565, + 0.00210663047619164, + 0.0024426004383713007, + 0.004324737470597029, + 0.002818052191287279, + 0.0015006341272965074, + 0.002880517626181245, + 0.004003400914371014, + 0.004339082632213831, + 0.00419255206361413, + 0.004222935531288385, + 0.004327498842030764, + 0.006087214685976505, + 0.004403532017022371, + 0.004944328684359789, + 0.005030530970543623, + 0.004913933575153351, + 0.007935888133943081, + 0.00835005659610033, + 0.0178199615329504, + 0.010096018202602863, + 0.006438197568058968, + 0.006808457896113396, + 0.009253568947315216, + 0.013476427644491196, + 0.008278700523078442, + 0.008929974399507046, + 0.007817897945642471, + 0.009524066932499409, + 0.00975219439715147, + 0.008095699362456799, + 0.006765047553926706, + 0.008647104725241661, + 0.0013853702694177628, + 0.002502142684534192, + 0.0019159583607688546, + 0.003893286222591996, + 0.0013157891808077693, + 0.0012863227166235447, + 0.002665088512003422, + 0.002260914770886302, + 0.005164375528693199, + 0.003065205179154873, + 0.010042024776339531, + 0.015132756903767586, + 0.0026537072844803333, + 0.002495046705007553, + 0.005948727019131184, + 0.007429381366819143, + 0.0037529501132667065, + 0.0035494393669068813, + 0.005067561753094196, + 0.003737666876986623, + 0.005560046527534723, + 0.005612609442323446, + 0.00635445024818182, + 0.0038534156046807766, + 0.006639929488301277, + 0.00407112343236804, + 0.006243958603590727, + 0.0031856107525527477, + 0.002228327328339219, + 0.0011688359081745148, + 0.0022259042598307133, + 0.00330187170766294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.899097740650177 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520751655101776 + ], + "opposite_logprob_tokens": [ + -1.5020751953125 + ], + "logprob_reward": -0.2520751655101776, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520751655101776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32521915435791016, + 0.10771961510181427, + 0.036304064095020294, + 0.08408939093351364, + 0.17762048542499542, + 0.041220784187316895, + 0.048787590116262436, + 0.08505065739154816, + 0.07508736103773117, + 0.018900873139500618 + ], + "credit_map": [ + 0.031011248007416725, + 0.04750993847846985, + 0.051489755511283875, + 0.03646795079112053, + 0.037683796137571335, + 0.047367263585329056, + 0.028355201706290245, + 0.045334022492170334, + 0.011949596926569939, + 0.011472723446786404, + 0.022323330864310265, + 0.011617113836109638, + 0.013068746775388718, + 0.011157006025314331, + 0.011899414472281933, + 0.014231677167117596, + 0.003968086093664169, + 0.0033490650821477175, + 0.0033374198246747255, + 0.0035927691496908665, + 0.005211047828197479, + 0.007014578208327293, + 0.004911199677735567, + 0.004919898696243763, + 0.006072745658457279, + 0.007759732659906149, + 0.01303272694349289, + 0.014999061822891235, + 0.008234905079007149, + 0.011415759101510048, + 0.010412040166556835, + 0.012162420898675919, + 0.02013378217816353, + 0.02362569235265255, + 0.019106386229395866, + 0.027164215222001076, + 0.025672966614365578, + 0.02075842022895813, + 0.019064869731664658, + 0.02209414355456829, + 0.004909907002002001, + 0.004526270087808371, + 0.005459269508719444, + 0.011402824893593788, + 0.003040208015590906, + 0.002821712289005518, + 0.0035353992134332657, + 0.005525193642824888, + 0.004344739485532045, + 0.0031448297668248415, + 0.0057445913553237915, + 0.01802384853363037, + 0.002659373451024294, + 0.0027705493848770857, + 0.004262029659003019, + 0.007837627083063126, + 0.008164840750396252, + 0.00984831154346466, + 0.012594087049365044, + 0.01794050820171833, + 0.006953679956495762, + 0.007698120549321175, + 0.01021038182079792, + 0.011640732176601887, + 0.008431184105575085, + 0.007282387465238571, + 0.01087088044732809, + 0.008895941078662872, + 0.00875070784240961, + 0.0077451984398067, + 0.015446067787706852, + 0.007664987817406654, + 0.002105931518599391, + 0.00215092278085649, + 0.002537676366046071, + 0.003973525483161211, + 0.0028170603327453136, + 0.0022136985789984465, + 0.0016720332205295563, + 0.0014300269540399313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.2520751655101776 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645151551812887 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645151551812887, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645151551812887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2289743274450302, + 0.3396604359149933, + 0.09289474785327911, + 0.03125016391277313, + 0.10312412679195404, + 0.04106197878718376, + 0.02920583076775074, + 0.06650646775960922, + 0.0379052497446537, + 0.029416721314191818 + ], + "credit_map": [ + 0.02303813025355339, + 0.032717783004045486, + 0.025375189259648323, + 0.02712981589138508, + 0.03189586102962494, + 0.03780646249651909, + 0.022935083135962486, + 0.028075985610485077, + 0.04132254421710968, + 0.029836326837539673, + 0.06785103678703308, + 0.03014262020587921, + 0.041470013558864594, + 0.038746923208236694, + 0.03717038780450821, + 0.05312058702111244, + 0.010064613074064255, + 0.011547152884304523, + 0.007756442297250032, + 0.009200037457048893, + 0.01038658618927002, + 0.010833396576344967, + 0.01878548040986061, + 0.014321038499474525, + 0.0030108054634183645, + 0.002981273690238595, + 0.0041959709487855434, + 0.004846211988478899, + 0.0031269469764083624, + 0.004984825849533081, + 0.0034961651545017958, + 0.0046079629100859165, + 0.012751693837344646, + 0.01238725520670414, + 0.014642934314906597, + 0.013081314042210579, + 0.013785813935101032, + 0.011045947670936584, + 0.011223568581044674, + 0.014205600135028362, + 0.005536900367587805, + 0.005374640226364136, + 0.004266364965587854, + 0.005010262597352266, + 0.00548425130546093, + 0.005032615270465612, + 0.005072347819805145, + 0.005284593440592289, + 0.002760461065918207, + 0.0028780237771570683, + 0.003573328722268343, + 0.010805453173816204, + 0.0011443844996392727, + 0.0012607432436197996, + 0.0023576649837195873, + 0.004425771068781614, + 0.006476766429841518, + 0.006682382430881262, + 0.01163590420037508, + 0.018748633563518524, + 0.0037387870252132416, + 0.003297125454992056, + 0.006420446094125509, + 0.009506420232355595, + 0.0038914703764021397, + 0.003616279922425747, + 0.0057395813055336475, + 0.0047244965098798275, + 0.004759257659316063, + 0.00395339448004961, + 0.007503910455852747, + 0.003716862527653575, + 0.0032003673259168863, + 0.0040170275606215, + 0.005348519887775183, + 0.005341206677258015, + 0.0023889143485575914, + 0.002645028755068779, + 0.0030498988926410675, + 0.0034257574006915092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.004645151551812887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977680366486311 + ], + "opposite_logprob_tokens": [ + -4.634776592254639 + ], + "logprob_reward": -0.00977680366486311, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977680366486311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.200544536113739, + 0.27669090032577515, + 0.03978132829070091, + 0.03187175095081329, + 0.23674388229846954, + 0.04357773810625076, + 0.02483600564301014, + 0.07695915549993515, + 0.052175864577293396, + 0.01681876927614212 + ], + "credit_map": [ + 0.019515251740813255, + 0.02684970386326313, + 0.02947726845741272, + 0.025264302268624306, + 0.022814178839325905, + 0.027264170348644257, + 0.020662013441324234, + 0.02869764342904091, + 0.033675458282232285, + 0.025160694494843483, + 0.05604838952422142, + 0.025032944977283478, + 0.033283475786447525, + 0.030201835557818413, + 0.029903719201683998, + 0.043384402990341187, + 0.0032728286460042, + 0.0033822334371507168, + 0.004809635225683451, + 0.0035679051652550697, + 0.005313202738761902, + 0.007051977328956127, + 0.006249895319342613, + 0.006133651826530695, + 0.004314408637583256, + 0.0025717183016240597, + 0.003540848381817341, + 0.0035615982487797737, + 0.005447648465633392, + 0.005380337592214346, + 0.0028908010572195053, + 0.004164388403296471, + 0.02908533252775669, + 0.030947139486670494, + 0.02605268359184265, + 0.03263291344046593, + 0.03193384408950806, + 0.026433134451508522, + 0.029783017933368683, + 0.02987581118941307, + 0.005863830912858248, + 0.0058039152063429356, + 0.003973317798227072, + 0.0053202686831355095, + 0.004948203451931477, + 0.004906137473881245, + 0.006985134910792112, + 0.0057769338600337505, + 0.0025109355337917805, + 0.002316152211278677, + 0.0029827815014868975, + 0.008760904893279076, + 0.0013980697840452194, + 0.0012824026634916663, + 0.0017930237809196115, + 0.0037917359732091427, + 0.007460954133421183, + 0.007914182730019093, + 0.012088362127542496, + 0.018756235018372536, + 0.005581515841186047, + 0.006073542404919863, + 0.008390791714191437, + 0.010693571530282497, + 0.005755707621574402, + 0.0057510496117174625, + 0.008152886293828487, + 0.006546196527779102, + 0.005729160737246275, + 0.005148896481841803, + 0.009672975167632103, + 0.005418993066996336, + 0.002757851965725422, + 0.001608084188774228, + 0.0033799230586737394, + 0.002795527223497629, + 0.0015294186305254698, + 0.001418055035173893, + 0.0017888516886159778, + 0.0015410572523251176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.00977680366486311 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thick, glowing plume of steam rising slowly from a natural hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, smooth river rocks lining the perimeter of the hot spring visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large, pale full moon visible ascending into a deep blue night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a classic granite stone lantern (*tōrō*) casting a warm, amber glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only perceptible movement the gentle swirls of steam and the soft flicker of the lantern flame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589079707860947 + ], + "opposite_logprob_tokens": [ + -3.8955891132354736 + ], + "logprob_reward": -0.020589079707860947, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589079707860947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2625340521335602, + 0.3537120521068573, + 0.04112408310174942, + 0.05516621470451355, + 0.06186608597636223, + 0.07100783288478851, + 0.030896931886672974, + 0.06489409506320953, + 0.035745300352573395, + 0.02305338717997074 + ], + "credit_map": [ + 0.025135517120361328, + 0.03337518870830536, + 0.04724351316690445, + 0.03542489558458328, + 0.025818951427936554, + 0.029798006638884544, + 0.02609066478908062, + 0.03964729234576225, + 0.044587720185518265, + 0.03170686960220337, + 0.07221195846796036, + 0.03172968700528145, + 0.042452968657016754, + 0.0371040403842926, + 0.038611236959695816, + 0.055307578295469284, + 0.003580904332920909, + 0.0036466869059950113, + 0.005845824722200632, + 0.005919590126723051, + 0.0038688473869115114, + 0.005498012527823448, + 0.006585524883121252, + 0.006178688257932663, + 0.0061835721135139465, + 0.005408796016126871, + 0.005162873305380344, + 0.0074239433743059635, + 0.007799462415277958, + 0.007897819392383099, + 0.007670717313885689, + 0.007619031704962254, + 0.004307039547711611, + 0.005261928308755159, + 0.012557804584503174, + 0.008999945595860481, + 0.005793808959424496, + 0.005414650775492191, + 0.010005554184317589, + 0.009525354951620102, + 0.0092596635222435, + 0.009056559763848782, + 0.007146099582314491, + 0.013096053153276443, + 0.009043249301612377, + 0.007614320609718561, + 0.007076866924762726, + 0.008715013042092323, + 0.0025871191173791885, + 0.0038848055992275476, + 0.003805107669904828, + 0.007195224054157734, + 0.002076334087178111, + 0.002223829971626401, + 0.004915731027722359, + 0.004208777099847794, + 0.007309941574931145, + 0.004793200641870499, + 0.0126609206199646, + 0.012973397970199585, + 0.0056814271956682205, + 0.005646362900733948, + 0.007932092994451523, + 0.007896753028035164, + 0.003758986946195364, + 0.0038654517848044634, + 0.006613488309085369, + 0.004284597001969814, + 0.004005464259535074, + 0.004439762327820063, + 0.0036164799239486456, + 0.005161067936569452, + 0.00224996916949749, + 0.0033027457538992167, + 0.004466538783162832, + 0.003953330684453249, + 0.002189709572121501, + 0.0015387671301141381, + 0.0026512539479881525, + 0.0027010717894881964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.020589079707860947 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_1/index_tensor([6424, 6424])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_1/index_tensor([6424, 6424])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8f2dd0e2933c20d8c4d807a3133d24b2b58b2b02 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_1/index_tensor([6424, 6424])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-speed tracking shot, the camera tilting sharply downward, closely follows a pitch-black crow. In an instant, the crow decisively folds its broad wings, tucking them tightly against its sides. Its inky black feathers gleam with a faint metallic sheen in the dim light. Its sturdy beak clenched tightly, its body tensed into a perfectly streamlined form. Then, like a flash of black lightning, it plunges vertically and unhesitatingly toward the blurred ground. A few loose feathers, ripped from its body by the gust of wind, tumble helplessly through the air. Against the backdrop, swirling, leaden storm clouds nearly engulf the sky, only faint, oppressive light glimmering through the gaps in the cloud cover. The entire image is imbued with the grainy, high-contrast quality of vintage film, its tones strictly constrained between varying shades of gray and abject black, creating an atmosphere of tension, coldness, and foreboding doom.", + "seed": 29909, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_1/index_6424_seed_29909_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13330671349540352, + "vlm_logprob": -0.13330671349540352, + "vlm_yes_no_margin": 3.449999952316284, + "vlm_token_credit": -0.13330671349540352, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5835265517234802 + ], + "opposite_logprob_tokens": [ + -0.8335265517234802 + ], + "logprob_reward": -0.5835265517234802, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5835265517234802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3225179612636566, + 0.0492832288146019, + 0.03841918334364891, + 0.042112451046705246, + 0.04190079867839813, + 0.038642365485429764, + 0.09291944652795792, + 0.06579197198152542, + 0.1288815587759018, + 0.1795310229063034 + ], + "credit_map": [ + 0.027356525883078575, + 0.1152624562382698, + 0.08628379553556442, + 0.016888394951820374, + 0.02294531650841236, + 0.016489306464791298, + 0.02602981962263584, + 0.011262382380664349, + 0.003978255204856396, + 0.005528471432626247, + 0.005036069545894861, + 0.004809912759810686, + 0.013443296775221825, + 0.007738180924206972, + 0.0045603192411363125, + 0.0041887275874614716, + 0.00620315782725811, + 0.004571438301354647, + 0.004182562697678804, + 0.002803065348416567, + 0.00658075837418437, + 0.00402654567733407, + 0.00379005610011518, + 0.006261602975428104, + 0.004506805445998907, + 0.004498521331697702, + 0.008083513006567955, + 0.0029561531264334917, + 0.0072121769189834595, + 0.006180069409310818, + 0.004012253601104021, + 0.004662962164729834, + 0.0061079589650034904, + 0.004205669276416302, + 0.005714481696486473, + 0.005323908757418394, + 0.006492896471172571, + 0.005610251799225807, + 0.006101755890995264, + 0.0023438797798007727, + 0.002928531263023615, + 0.002465009456500411, + 0.004586041904985905, + 0.010619000531733036, + 0.0030635560397058725, + 0.0072234878316521645, + 0.004027259070426226, + 0.0037294807843863964, + 0.0089297229424119, + 0.007283767219632864, + 0.026826607063412666, + 0.010418921709060669, + 0.005918972194194794, + 0.01754642091691494, + 0.0070316675119102, + 0.008963373489677906, + 0.005980708170682192, + 0.005738705396652222, + 0.012589395977556705, + 0.013190971687436104, + 0.004435186740010977, + 0.007765675894916058, + 0.00821000337600708, + 0.00788133405148983, + 0.010163393802940845, + 0.006313189398497343, + 0.02409791573882103, + 0.04385044425725937, + 0.0055921729654073715, + 0.006610511802136898, + 0.01528695784509182, + 0.01696699485182762, + 0.00880960188806057, + 0.024505257606506348, + 0.04958212003111839, + 0.03420209512114525, + 0.013356655836105347, + 0.009616493247449398, + 0.02226642705500126, + 0.017192400991916656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.5835265517234802 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255736779421568 + ], + "opposite_logprob_tokens": [ + -4.387557506561279 + ], + "logprob_reward": -0.01255736779421568, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255736779421568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1161644384264946, + 0.13896909356117249, + 0.2281990498304367, + 0.050014838576316833, + 0.050473231822252274, + 0.029241053387522697, + 0.04255000129342079, + 0.024812452495098114, + 0.2415246069431305, + 0.07805120944976807 + ], + "credit_map": [ + 0.011129515245556831, + 0.025216907262802124, + 0.019811291247606277, + 0.009858605451881886, + 0.01942109316587448, + 0.011010194197297096, + 0.012366563081741333, + 0.007350277621299028, + 0.009794922545552254, + 0.015346651896834373, + 0.012741073966026306, + 0.012605952098965645, + 0.03893677890300751, + 0.025472354143857956, + 0.013787142932415009, + 0.010284225456416607, + 0.03704489395022392, + 0.024375613778829575, + 0.038729477673769, + 0.019668972119688988, + 0.032305482774972916, + 0.018624821677803993, + 0.02264946699142456, + 0.034800346940755844, + 0.006110487040132284, + 0.004476544912904501, + 0.009082975797355175, + 0.003054380416870117, + 0.008705167099833488, + 0.007325268350541592, + 0.005117439199239016, + 0.006142579484730959, + 0.005904889199882746, + 0.004493163898587227, + 0.00809350237250328, + 0.004894808866083622, + 0.0061745899729430676, + 0.007967670448124409, + 0.007686284836381674, + 0.005258327350020409, + 0.0027952403761446476, + 0.002120537217706442, + 0.004973824135959148, + 0.006019955035299063, + 0.001994799356907606, + 0.004499961622059345, + 0.0036783781833946705, + 0.003158360021188855, + 0.004115207586437464, + 0.002996290335431695, + 0.012809552252292633, + 0.004793757572770119, + 0.0029996444936841726, + 0.008319241926074028, + 0.0028848154470324516, + 0.0036314944736659527, + 0.0020478873047977686, + 0.002415861003100872, + 0.003958086483180523, + 0.004216710105538368, + 0.002003708854317665, + 0.003168909577652812, + 0.0023534458596259356, + 0.004647844471037388, + 0.007824473083019257, + 0.009131635539233685, + 0.052316147834062576, + 0.0997495949268341, + 0.00747050391510129, + 0.00955699197947979, + 0.016620852053165436, + 0.03885442763566971, + 0.006492037791758776, + 0.007578443735837936, + 0.011880788020789623, + 0.010104133747518063, + 0.008440053090453148, + 0.010188129730522633, + 0.015204748138785362, + 0.008162874728441238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.01255736779421568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865364357829094 + ], + "opposite_logprob_tokens": [ + -3.0486536026000977 + ], + "logprob_reward": -0.04865364357829094, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865364357829094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2649284601211548, + 0.03541722148656845, + 0.14759233593940735, + 0.05728072300553322, + 0.03239951655268669, + 0.07060330361127853, + 0.04085348919034004, + 0.04766397178173065, + 0.04867219179868698, + 0.2545888125896454 + ], + "credit_map": [ + 0.023121675476431847, + 0.10067173093557358, + 0.07035636901855469, + 0.014633622020483017, + 0.015999265015125275, + 0.007982674986124039, + 0.023874469101428986, + 0.008288661018013954, + 0.003007132327184081, + 0.004582773428410292, + 0.003670269390568137, + 0.0028828722424805164, + 0.009534254670143127, + 0.005466500297188759, + 0.0033902577124536037, + 0.0028831602539867163, + 0.028490636497735977, + 0.015184229239821434, + 0.027360031381249428, + 0.012982631102204323, + 0.016934597864747047, + 0.010144640691578388, + 0.014468731358647346, + 0.022026842460036278, + 0.0045035541988909245, + 0.009024607948958874, + 0.00574946915730834, + 0.004147137049585581, + 0.012526516802608967, + 0.012038187123835087, + 0.005484673660248518, + 0.003806576831266284, + 0.0027484423480927944, + 0.0030981749296188354, + 0.0032375785522162914, + 0.005055632907897234, + 0.00485624372959137, + 0.004703101702034473, + 0.00599629757925868, + 0.002704047132283449, + 0.00703056063503027, + 0.005275258794426918, + 0.009387261234223843, + 0.009442669339478016, + 0.011308576911687851, + 0.011229078285396099, + 0.009488223120570183, + 0.007441674824804068, + 0.0070032766088843346, + 0.0026767533272504807, + 0.0077358693815767765, + 0.004187203478068113, + 0.003854616777971387, + 0.007511764299124479, + 0.0034978208132088184, + 0.004386186134070158, + 0.004382682964205742, + 0.003721118438988924, + 0.007857786491513252, + 0.00982903502881527, + 0.0037438184954226017, + 0.006688443943858147, + 0.0045178732834756374, + 0.006923210807144642, + 0.004794503562152386, + 0.004107865039259195, + 0.007038576994091272, + 0.007386920973658562, + 0.004984819330275059, + 0.0052437433041632175, + 0.009215728379786015, + 0.005900034215301275, + 0.012899375520646572, + 0.03293788433074951, + 0.0476987324655056, + 0.06810230016708374, + 0.018438933417201042, + 0.018815286457538605, + 0.03166871517896652, + 0.024027561768889427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.04865364357829094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818203739821911 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.01818203739821911, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818203739821911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1611153483390808, + 0.04688257724046707, + 0.016193952411413193, + 0.04202713817358017, + 0.018374571576714516, + 0.042003799229860306, + 0.03750193491578102, + 0.08770692348480225, + 0.48141124844551086, + 0.06678248196840286 + ], + "credit_map": [ + 0.01376374252140522, + 0.06171046197414398, + 0.0426027849316597, + 0.008601856417953968, + 0.010811113752424717, + 0.004755124915391207, + 0.01454917062073946, + 0.004321093671023846, + 0.003795457538217306, + 0.005147173535078764, + 0.004993306938558817, + 0.004323478788137436, + 0.015080844052135944, + 0.00591388950124383, + 0.004206663463264704, + 0.003421766683459282, + 0.0026098182424902916, + 0.0017882711254060268, + 0.00253633176907897, + 0.0011931429617106915, + 0.002248914912343025, + 0.0020326257217675447, + 0.0016414817655459046, + 0.0021433664951473475, + 0.0029144275467842817, + 0.0052160294726490974, + 0.005167854949831963, + 0.003342968877404928, + 0.00781348068267107, + 0.007721385452896357, + 0.005022092256695032, + 0.0048288991674780846, + 0.0017660933081060648, + 0.0016556092305108905, + 0.002297046361491084, + 0.0025768019258975983, + 0.00263652834109962, + 0.0032609773334115744, + 0.002465610858052969, + 0.001715903403237462, + 0.004102002363651991, + 0.003925337456166744, + 0.006481199525296688, + 0.00815222505480051, + 0.0041986918076872826, + 0.006277047097682953, + 0.00539613189175725, + 0.003471165196970105, + 0.0037499123718589544, + 0.0031329155899584293, + 0.007382568903267384, + 0.0049674189649522305, + 0.0035661582369357347, + 0.006307638715952635, + 0.004685486666858196, + 0.003709836397320032, + 0.008225406520068645, + 0.00926155410706997, + 0.014167653396725655, + 0.014937081374228, + 0.00768401799723506, + 0.01293693296611309, + 0.007110849488526583, + 0.013383423909544945, + 0.018981851637363434, + 0.024204671382904053, + 0.09815871715545654, + 0.1821034550666809, + 0.02663789875805378, + 0.025433380156755447, + 0.03630571812391281, + 0.06958555430173874, + 0.0036951641086488962, + 0.011062427423894405, + 0.013566836714744568, + 0.011116706766188145, + 0.0060840691439807415, + 0.005548672750592232, + 0.008621575310826302, + 0.007087029982358217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.01818203739821911 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036139669828116894 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036139669828116894, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036139669828116894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.311598539352417, + 0.05192206799983978, + 0.022546891123056412, + 0.0241236574947834, + 0.031553998589515686, + 0.02949383296072483, + 0.018654000014066696, + 0.04933138191699982, + 0.037425171583890915, + 0.4233504831790924 + ], + "credit_map": [ + 0.02669537626206875, + 0.11955325305461884, + 0.07830287516117096, + 0.017326246947050095, + 0.019186800345778465, + 0.010032582096755505, + 0.02997066080570221, + 0.010530714876949787, + 0.003496967488899827, + 0.005615492817014456, + 0.005659207701683044, + 0.004731697496026754, + 0.017375996336340904, + 0.006294679827988148, + 0.004727731458842754, + 0.004020290914922953, + 0.0020163182634860277, + 0.002634629840031266, + 0.002971293404698372, + 0.0025303608272224665, + 0.004570262506604195, + 0.002927771070972085, + 0.002286272356286645, + 0.002609981456771493, + 0.0018775942735373974, + 0.002129546133801341, + 0.00559338228777051, + 0.0014068027958273888, + 0.0032367678359150887, + 0.0065264091826975346, + 0.001476183533668518, + 0.0018769724993035197, + 0.0036040216218680143, + 0.0034656559582799673, + 0.004943091422319412, + 0.003607800230383873, + 0.003699918044731021, + 0.004978999029844999, + 0.004369096364825964, + 0.0028854170814156532, + 0.003018449293449521, + 0.002787394914776087, + 0.003920463379472494, + 0.004319777246564627, + 0.004164078738540411, + 0.004417503252625465, + 0.003807676024734974, + 0.003058488480746746, + 0.003163657383993268, + 0.001660345820710063, + 0.0030638715252280235, + 0.002202011877670884, + 0.001292217755690217, + 0.004023965448141098, + 0.0014307982055470347, + 0.0018171326955780387, + 0.004868578165769577, + 0.006053137127310038, + 0.009613723494112492, + 0.007406558841466904, + 0.004767393693327904, + 0.006624220870435238, + 0.0041588833555579185, + 0.00583888404071331, + 0.002329924376681447, + 0.002535642124712467, + 0.00890071876347065, + 0.012257742695510387, + 0.0022072582505643368, + 0.002078880090266466, + 0.003543747588992119, + 0.003571254899725318, + 0.023142069578170776, + 0.05604392662644386, + 0.0857943519949913, + 0.10836941003799438, + 0.028643282130360603, + 0.025870652869343758, + 0.05314535275101662, + 0.04234141856431961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0036139669828116894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed tracking shot, the camera tilting sharply downward, closely follows a pitch-black crow. In an instant, the crow decisively folds its broad wings, tucking them tightly against its sides. Its inky black feathers gleam with a faint metallic sheen in the dim light. Its sturdy beak clenched tightly, its body tensed into a perfectly streamlined form. Then, like a flash of black lightning, it plunges vertically and unhesitatingly toward the blurred ground. A few loose feathers, ripped from its body by the gust of wind, tumble helplessly through the air. Against the backdrop, swirling, leaden storm clouds nearly engulf the sky, only faint, oppressive light glimmering through the gaps in the cloud cover. The entire image is imbued with the grainy, high-contrast quality of vintage film, its tones strictly constrained between varying shades of gray and abject black, creating an atmosphere of tension, coldness, and foreboding doom.", + "seed": 29910, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_1/index_6424_seed_29910_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1965852172113955, + "vlm_logprob": -0.1965852172113955, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.1965852172113955, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8402578830718994 + ], + "opposite_logprob_tokens": [ + -0.5902578830718994 + ], + "logprob_reward": -0.8402578830718994, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8402578830718994, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07050818204879761, + 0.027867425233125687, + 0.09102403372526169, + 0.051085472106933594, + 0.04795713722705841, + 0.29188117384910583, + 0.06477799266576767, + 0.21133700013160706, + 0.07161226868629456, + 0.07194926589727402 + ], + "credit_map": [ + 0.009345388039946556, + 0.010383278131484985, + 0.010849954560399055, + 0.008408777415752411, + 0.00771563732996583, + 0.011970369145274162, + 0.006091439630836248, + 0.005743339657783508, + 0.002611994743347168, + 0.0034048159141093493, + 0.004473124630749226, + 0.00396311841905117, + 0.003207437926903367, + 0.0045158532448112965, + 0.002971796551719308, + 0.0027192847337573767, + 0.009249900467693806, + 0.007233199663460255, + 0.014633671380579472, + 0.00651025352999568, + 0.008557179011404514, + 0.02435871958732605, + 0.010482829995453358, + 0.009998287074267864, + 0.00435201870277524, + 0.005958879832178354, + 0.010256451554596424, + 0.006235514301806688, + 0.006287463940680027, + 0.005718922708183527, + 0.005347895435988903, + 0.006928327027708292, + 0.0032122828997671604, + 0.004657694138586521, + 0.00721460347995162, + 0.004338120575994253, + 0.008636489510536194, + 0.0075425575487315655, + 0.005938218906521797, + 0.0064171720296144485, + 0.03062657080590725, + 0.024812353774905205, + 0.053743764758110046, + 0.02666516602039337, + 0.030933933332562447, + 0.06485815346240997, + 0.021398484706878662, + 0.03884274885058403, + 0.007172415964305401, + 0.008441885933279991, + 0.01048237830400467, + 0.00724587868899107, + 0.01019426342099905, + 0.006025020033121109, + 0.0068470085971057415, + 0.008369140326976776, + 0.019005954265594482, + 0.03629772365093231, + 0.032093800604343414, + 0.02264506183564663, + 0.021117182448506355, + 0.029672132804989815, + 0.02420090325176716, + 0.026304254308342934, + 0.005544010084122419, + 0.007644006051123142, + 0.014377144165337086, + 0.008166573941707611, + 0.008745454251766205, + 0.013442360796034336, + 0.007106316275894642, + 0.006586403585970402, + 0.0036292406730353832, + 0.005826330278068781, + 0.024589544162154198, + 0.004193002358078957, + 0.025507427752017975, + 0.0028626855928450823, + 0.002878808416426182, + 0.00246222410351038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.8402578830718994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003026455640793 + ], + "opposite_logprob_tokens": [ + -2.6950302124023438 + ], + "logprob_reward": -0.07003026455640793, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003026455640793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0762193575501442, + 0.03880797326564789, + 0.060737013816833496, + 0.4756854474544525, + 0.127615824341774, + 0.07964137941598892, + 0.03698520362377167, + 0.042718321084976196, + 0.01627783663570881, + 0.045311637222766876 + ], + "credit_map": [ + 0.012354444712400436, + 0.009350501000881195, + 0.010842740535736084, + 0.00982952956110239, + 0.011785417795181274, + 0.009894193150103092, + 0.007823066785931587, + 0.004339470528066158, + 0.0034567643888294697, + 0.004179936368018389, + 0.005858667194843292, + 0.004601988475769758, + 0.004996255040168762, + 0.006052715238183737, + 0.005211760755628347, + 0.004449891857802868, + 0.007267499342560768, + 0.0065546296536922455, + 0.007089465390890837, + 0.0062187230214476585, + 0.006650439463555813, + 0.0118368249386549, + 0.008735607378184795, + 0.006383826490491629, + 0.03927292674779892, + 0.040471695363521576, + 0.08442560583353043, + 0.05099811777472496, + 0.061397433280944824, + 0.07197168469429016, + 0.05701436474919319, + 0.07013363391160965, + 0.011924656108021736, + 0.010226340964436531, + 0.02166777104139328, + 0.010627103969454765, + 0.013440050184726715, + 0.019620856270194054, + 0.020001767203211784, + 0.02010728046298027, + 0.010669317096471786, + 0.008602616377174854, + 0.013893771916627884, + 0.007772062439471483, + 0.0070866928435862064, + 0.01611902005970478, + 0.005891564302146435, + 0.00960633810609579, + 0.0033345234114676714, + 0.003419485641643405, + 0.008878384716808796, + 0.004128039814531803, + 0.004056780599057674, + 0.0045509859919548035, + 0.002720582764595747, + 0.005896422080695629, + 0.0029218175914138556, + 0.010950087569653988, + 0.009590360336005688, + 0.004242239985615015, + 0.003240112913772464, + 0.004624367691576481, + 0.0034972133580595255, + 0.003652128390967846, + 0.0015332865295931697, + 0.0031861846800893545, + 0.002497702604159713, + 0.001982445362955332, + 0.0021297072526067495, + 0.0020558955147862434, + 0.0013738626148551702, + 0.0015187531244009733, + 0.002086975146085024, + 0.004529349505901337, + 0.010904906317591667, + 0.0033658938482403755, + 0.015959490090608597, + 0.0030067365150898695, + 0.0038693107198923826, + 0.0015889795031398535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.07003026455640793 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369683399796486 + ], + "opposite_logprob_tokens": [ + -3.4086968898773193 + ], + "logprob_reward": -0.03369683399796486, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369683399796486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034509576857089996, + 0.021476389840245247, + 0.051532208919525146, + 0.24540720880031586, + 0.11965109407901764, + 0.31898608803749084, + 0.031934455037117004, + 0.043566495180130005, + 0.08010304719209671, + 0.05283340811729431 + ], + "credit_map": [ + 0.0044321222230792046, + 0.005069410894066095, + 0.003838628064841032, + 0.005334292072802782, + 0.004971882328391075, + 0.004330714698880911, + 0.003755070036277175, + 0.0027774563059210777, + 0.0019547578413039446, + 0.002790736733004451, + 0.003191924188286066, + 0.002552888123318553, + 0.0030275953467935324, + 0.003370746038854122, + 0.0024378651287406683, + 0.002149876905605197, + 0.005798801314085722, + 0.004733012057840824, + 0.008248211815953255, + 0.003899590577930212, + 0.0051443870179355145, + 0.012352809309959412, + 0.005884923040866852, + 0.005470470059663057, + 0.021527597680687904, + 0.021191032603383064, + 0.04028068110346794, + 0.02564322203397751, + 0.03203022480010986, + 0.0376754105091095, + 0.029777104035019875, + 0.03728194162249565, + 0.007316660601645708, + 0.009644473902881145, + 0.016810867935419083, + 0.01201990433037281, + 0.022168438881635666, + 0.01991935819387436, + 0.016823329031467438, + 0.014948050491511822, + 0.04736853763461113, + 0.04263770207762718, + 0.055274542421102524, + 0.021279556676745415, + 0.02919895015656948, + 0.0647350400686264, + 0.021272096782922745, + 0.037219662219285965, + 0.0032861605286598206, + 0.0030855084769427776, + 0.006898729596287012, + 0.003819562727585435, + 0.0036474077496677637, + 0.0032798349857330322, + 0.0032240059226751328, + 0.004693247843533754, + 0.004529204685240984, + 0.005501285661011934, + 0.00502424780279398, + 0.005101295188069344, + 0.005064834374934435, + 0.006297008600085974, + 0.005758205894380808, + 0.006290416698902845, + 0.004242058377712965, + 0.0066459705121815205, + 0.01375492475926876, + 0.005474472418427467, + 0.01240166649222374, + 0.028320088982582092, + 0.005275937262922525, + 0.003987927921116352, + 0.002440226264297962, + 0.004166597966104746, + 0.016273509711027145, + 0.003316788235679269, + 0.020075393840670586, + 0.002433943562209606, + 0.002757203998044133, + 0.001369739999063313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.03369683399796486 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369395062327385 + ], + "opposite_logprob_tokens": [ + -3.40869402885437 + ], + "logprob_reward": -0.03369395062327385, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369395062327385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05412669852375984, + 0.022180210798978806, + 0.12434501945972443, + 0.05283763259649277, + 0.23581011593341827, + 0.154993936419487, + 0.07185035198926926, + 0.14681817591190338, + 0.05872306972742081, + 0.07831472158432007 + ], + "credit_map": [ + 0.008088893257081509, + 0.005736913066357374, + 0.007329252548515797, + 0.006834140047430992, + 0.0076004983857274055, + 0.008823312819004059, + 0.005180245731025934, + 0.004533444996923208, + 0.0017407059203833342, + 0.002482484793290496, + 0.0032108027953654528, + 0.0032179036643356085, + 0.0020843411330133677, + 0.002467315411195159, + 0.004204972181469202, + 0.0027716842014342546, + 0.012974337674677372, + 0.009473750367760658, + 0.01857142150402069, + 0.011888690292835236, + 0.010921929031610489, + 0.026010287925601006, + 0.020164301618933678, + 0.014340296387672424, + 0.004108187276870012, + 0.0051866257563233376, + 0.008748373948037624, + 0.004838089924305677, + 0.00666778814047575, + 0.008139862678945065, + 0.006899284664541483, + 0.008249422535300255, + 0.024752043187618256, + 0.017086034640669823, + 0.03471345826983452, + 0.018595613539218903, + 0.026939881965517998, + 0.04414035752415657, + 0.03902893885970116, + 0.030553793534636497, + 0.014991077594459057, + 0.013940495438873768, + 0.028624214231967926, + 0.02009718306362629, + 0.013029790483415127, + 0.029677903279662132, + 0.015598442405462265, + 0.01903483085334301, + 0.007025771774351597, + 0.00660701934248209, + 0.017803549766540527, + 0.009227783419191837, + 0.0056120422668755054, + 0.007152187172323465, + 0.007847188971936703, + 0.010574807412922382, + 0.013948515988886356, + 0.020607655867934227, + 0.01987597718834877, + 0.015582204796373844, + 0.015527486801147461, + 0.02342280186712742, + 0.0178877804428339, + 0.019965760409832, + 0.004266547039151192, + 0.008566255681216717, + 0.01044874545186758, + 0.007292953785508871, + 0.00760786235332489, + 0.009224322624504566, + 0.007160681299865246, + 0.004155697301030159, + 0.0038747095968574286, + 0.006801839452236891, + 0.020390570163726807, + 0.006195430643856525, + 0.026565009728074074, + 0.005032660439610481, + 0.006436372175812721, + 0.0030181240290403366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.03369395062327385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247153807431459 + ], + "opposite_logprob_tokens": [ + -5.255247116088867 + ], + "logprob_reward": -0.005247153807431459, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247153807431459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03961803764104843, + 0.018557190895080566, + 0.045918531715869904, + 0.04737411066889763, + 0.048035260289907455, + 0.03357750549912453, + 0.03588838875293732, + 0.18400786817073822, + 0.16014288365840912, + 0.3868802785873413 + ], + "credit_map": [ + 0.005279831122606993, + 0.0037591420114040375, + 0.007582528982311487, + 0.0038087975699454546, + 0.004407624248415232, + 0.005479224491864443, + 0.003821836318820715, + 0.0054790545254945755, + 0.0016545284306630492, + 0.0014518783427774906, + 0.0028553290758281946, + 0.0020526943262666464, + 0.001777445082552731, + 0.0022652947809547186, + 0.0031908995006233454, + 0.003309123683720827, + 0.0041461531072855, + 0.0029292567633092403, + 0.007803381886333227, + 0.0037290877662599087, + 0.0035609756596386433, + 0.012332494370639324, + 0.006156901828944683, + 0.005260278005152941, + 0.004089272115379572, + 0.004417384043335915, + 0.007879812270402908, + 0.004920531529933214, + 0.0060844579711556435, + 0.007231282070279121, + 0.005797315388917923, + 0.006954053416848183, + 0.00609222985804081, + 0.00302868545986712, + 0.007418879307806492, + 0.0038874438032507896, + 0.004962384235113859, + 0.006945719476789236, + 0.007387725170701742, + 0.008312191814184189, + 0.004505758173763752, + 0.005933756940066814, + 0.0039549157954752445, + 0.004291042685508728, + 0.0048802997916936874, + 0.003270805114880204, + 0.004245108459144831, + 0.0024958201684057713, + 0.0032858061604201794, + 0.00464056758210063, + 0.008049547672271729, + 0.003506916807964444, + 0.00424441322684288, + 0.0034661772660911083, + 0.0030438369140028954, + 0.005651124753057957, + 0.014355759136378765, + 0.03770776838064194, + 0.04562080651521683, + 0.02039213478565216, + 0.013955177739262581, + 0.015054461546242237, + 0.016125639900565147, + 0.02079610526561737, + 0.009769072756171227, + 0.01882767118513584, + 0.019670406356453896, + 0.01601584441959858, + 0.024309808388352394, + 0.04487040638923645, + 0.014573376625776291, + 0.01210630964487791, + 0.017181307077407837, + 0.03672635555267334, + 0.10201642662286758, + 0.02707577869296074, + 0.1478007286787033, + 0.021581731736660004, + 0.02165340818464756, + 0.012844531796872616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.005247153807431459 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed tracking shot, the camera tilting sharply downward, closely follows a pitch-black crow. In an instant, the crow decisively folds its broad wings, tucking them tightly against its sides. Its inky black feathers gleam with a faint metallic sheen in the dim light. Its sturdy beak clenched tightly, its body tensed into a perfectly streamlined form. Then, like a flash of black lightning, it plunges vertically and unhesitatingly toward the blurred ground. A few loose feathers, ripped from its body by the gust of wind, tumble helplessly through the air. Against the backdrop, swirling, leaden storm clouds nearly engulf the sky, only faint, oppressive light glimmering through the gaps in the cloud cover. The entire image is imbued with the grainy, high-contrast quality of vintage film, its tones strictly constrained between varying shades of gray and abject black, creating an atmosphere of tension, coldness, and foreboding doom.", + "seed": 29911, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_1/index_6424_seed_29911_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2590626263990998, + "vlm_logprob": -0.2590626263990998, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.2590626263990998, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9831259250640869 + ], + "opposite_logprob_tokens": [ + -0.4831259250640869 + ], + "logprob_reward": -0.9831259250640869, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9831259250640869, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.053158607333898544, + 0.10623805969953537, + 0.02429773099720478, + 0.09534051269292831, + 0.0652245283126831, + 0.049258098006248474, + 0.039888329803943634, + 0.035487718880176544, + 0.48264560103416443, + 0.04846077412366867 + ], + "credit_map": [ + 0.00843209307640791, + 0.0089925117790699, + 0.006129447836428881, + 0.004688078537583351, + 0.010463242419064045, + 0.0055901240557432175, + 0.004343983251601458, + 0.004519127309322357, + 0.012253058142960072, + 0.015942927449941635, + 0.009639873169362545, + 0.012892123311758041, + 0.018542874604463577, + 0.01701813004910946, + 0.010058479383587837, + 0.009890594519674778, + 0.004046870395541191, + 0.0034096415620297194, + 0.002697189338505268, + 0.0028410281520336866, + 0.00269912788644433, + 0.0034948831889778376, + 0.0025941163767129183, + 0.0025148745626211166, + 0.009217645972967148, + 0.00813307799398899, + 0.020723139867186546, + 0.009051218628883362, + 0.008227278478443623, + 0.012915153056383133, + 0.006704170256853104, + 0.02036883495748043, + 0.004066566005349159, + 0.005738354288041592, + 0.00746966851875186, + 0.008330161683261395, + 0.011530082672834396, + 0.014075319282710552, + 0.008338520303368568, + 0.005675856489688158, + 0.004129158798605204, + 0.0033004540018737316, + 0.007397063076496124, + 0.0101931132376194, + 0.007432310376316309, + 0.005484536290168762, + 0.005403243936598301, + 0.005918217357248068, + 0.004367802757769823, + 0.003500324208289385, + 0.007645834237337112, + 0.005119842942804098, + 0.004633407574146986, + 0.0043127029202878475, + 0.0036722272634506226, + 0.006636190693825483, + 0.005197422578930855, + 0.004852559417486191, + 0.006433276925235987, + 0.0039611076936125755, + 0.003457289654761553, + 0.0035781438928097486, + 0.0028116872999817133, + 0.0051962328143417835, + 0.02303931675851345, + 0.19178380072116852, + 0.14574548602104187, + 0.02194306254386902, + 0.05389812961220741, + 0.02235168032348156, + 0.010519804432988167, + 0.01336435042321682, + 0.004931146278977394, + 0.006315106060355902, + 0.003640988143160939, + 0.004374289885163307, + 0.0073593673296272755, + 0.015441512688994408, + 0.003072477178648114, + 0.0033258874900639057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.9831259250640869 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003515213727951 + ], + "opposite_logprob_tokens": [ + -2.695035219192505 + ], + "logprob_reward": -0.07003515213727951, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003515213727951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1372750848531723, + 0.06125275790691376, + 0.11264754831790924, + 0.2055472731590271, + 0.0985870286822319, + 0.05623926594853401, + 0.06093314662575722, + 0.040874071419239044, + 0.17917867004871368, + 0.04746516793966293 + ], + "credit_map": [ + 0.02028750814497471, + 0.01942615956068039, + 0.010643425397574902, + 0.015018914826214314, + 0.012912717647850513, + 0.013741003349423409, + 0.01934637874364853, + 0.025898978114128113, + 0.009662195108830929, + 0.007779135834425688, + 0.006225847639143467, + 0.005154372658580542, + 0.009826735593378544, + 0.008513488806784153, + 0.004810389131307602, + 0.009280593134462833, + 0.012514645233750343, + 0.013156836852431297, + 0.011193528771400452, + 0.0121117252856493, + 0.01669578067958355, + 0.020644083619117737, + 0.016011152416467667, + 0.010319792665541172, + 0.017927195876836777, + 0.017995398491621017, + 0.04512869194149971, + 0.01730755902826786, + 0.015716778114438057, + 0.026675939559936523, + 0.01966848038136959, + 0.045127224177122116, + 0.0062301019206643105, + 0.010035239160060883, + 0.012751298025250435, + 0.013354259543120861, + 0.018167847767472267, + 0.022372756153345108, + 0.008755868300795555, + 0.006919661071151495, + 0.004920934326946735, + 0.0043829320929944515, + 0.007304077036678791, + 0.010730436071753502, + 0.011004626750946045, + 0.006402673665434122, + 0.006564852315932512, + 0.004928733222186565, + 0.0076860059052705765, + 0.005182578228414059, + 0.010152680799365044, + 0.006825585383921862, + 0.0057639773003757, + 0.007744552567601204, + 0.005900904536247253, + 0.01167686190456152, + 0.0034860281739383936, + 0.005349511746317148, + 0.005745221395045519, + 0.0043945894576609135, + 0.005571465473622084, + 0.00710718659684062, + 0.004978799726814032, + 0.004241272807121277, + 0.01267816498875618, + 0.05753401294350624, + 0.0636994019150734, + 0.009131046012043953, + 0.010321726091206074, + 0.00856397021561861, + 0.007098807021975517, + 0.010151532478630543, + 0.0030033704824745655, + 0.008405013009905815, + 0.006125129759311676, + 0.005507830996066332, + 0.006995664909482002, + 0.006897513754665852, + 0.0050188833847641945, + 0.005511762108653784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.07003515213727951 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036678850650787 + ], + "opposite_logprob_tokens": [ + -1.9103667736053467 + ], + "logprob_reward": -0.16036678850650787, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036678850650787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11879309266805649, + 0.1252087503671646, + 0.06865692138671875, + 0.14785346388816833, + 0.08489584177732468, + 0.09708335995674133, + 0.0764019787311554, + 0.08847891539335251, + 0.05593240633606911, + 0.13669535517692566 + ], + "credit_map": [ + 0.015290996059775352, + 0.015983030200004578, + 0.011640365235507488, + 0.012139552272856236, + 0.028046539053320885, + 0.016453348100185394, + 0.008924837224185467, + 0.010314421728253365, + 0.015206883661448956, + 0.019347159191966057, + 0.011454329825937748, + 0.014997327700257301, + 0.022996826097369194, + 0.020374085754156113, + 0.010471595451235771, + 0.01036052219569683, + 0.008188561536371708, + 0.008703420870006084, + 0.009384002536535263, + 0.014355979859828949, + 0.007335744332522154, + 0.007037350907921791, + 0.007241335231810808, + 0.006410518195480108, + 0.014457100071012974, + 0.014828619547188282, + 0.032636720687150955, + 0.014602495357394218, + 0.01202812697738409, + 0.014862174168229103, + 0.011025548912584782, + 0.03341268002986908, + 0.0052499789744615555, + 0.009579473175108433, + 0.008833943866193295, + 0.010336469858884811, + 0.015507997013628483, + 0.018541714176535606, + 0.010791775770485401, + 0.006054483354091644, + 0.008511552587151527, + 0.00690066022798419, + 0.01093085017055273, + 0.016355998814105988, + 0.013706707395613194, + 0.011382710188627243, + 0.01468680426478386, + 0.014608071185648441, + 0.0075891283340752125, + 0.00723255192860961, + 0.011176098138093948, + 0.009253060445189476, + 0.008396252058446407, + 0.010500920936465263, + 0.008609866723418236, + 0.01364409364759922, + 0.007537831086665392, + 0.01314567495137453, + 0.009857820346951485, + 0.008533535525202751, + 0.016056593507528305, + 0.020453426986932755, + 0.006033245008438826, + 0.006860785186290741, + 0.00757400318980217, + 0.008215460926294327, + 0.0105426711961627, + 0.006290716119110584, + 0.00850122794508934, + 0.006370584946125746, + 0.003304904792457819, + 0.005132828839123249, + 0.013763772323727608, + 0.018761038780212402, + 0.012107138521969318, + 0.011420675553381443, + 0.02107413485646248, + 0.0431264266371727, + 0.00835962500423193, + 0.008082537911832333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.16036678850650787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789700523018837 + ], + "opposite_logprob_tokens": [ + -2.578969955444336 + ], + "logprob_reward": -0.0789700523018837, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789700523018837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05129214748740196, + 0.2100210040807724, + 0.09341471642255783, + 0.2588152587413788, + 0.04908428713679314, + 0.07391369342803955, + 0.04439666122198105, + 0.07985865324735641, + 0.08769046515226364, + 0.051513101905584335 + ], + "credit_map": [ + 0.009059217758476734, + 0.007625995669513941, + 0.004518229980021715, + 0.0037937976885586977, + 0.012556043453514576, + 0.005152078345417976, + 0.0035289591178297997, + 0.005057831294834614, + 0.02426866441965103, + 0.030672546476125717, + 0.0203260388225317, + 0.025102609768509865, + 0.03789201378822327, + 0.03210015594959259, + 0.02283308655023575, + 0.01682586781680584, + 0.008596059866249561, + 0.012905890122056007, + 0.012327025644481182, + 0.013339466415345669, + 0.01266072504222393, + 0.012980368919670582, + 0.011486855335533619, + 0.009118324145674706, + 0.02632587030529976, + 0.01780305616557598, + 0.06189790368080139, + 0.020287441089749336, + 0.018555479124188423, + 0.03541675582528114, + 0.016029400750994682, + 0.06249937415122986, + 0.002859331900253892, + 0.006446906365454197, + 0.004411875735968351, + 0.005355497822165489, + 0.009864495135843754, + 0.011713539250195026, + 0.00500755337998271, + 0.0034250831231474876, + 0.006364284548908472, + 0.005540972109884024, + 0.006933700293302536, + 0.00816353689879179, + 0.016054667532444, + 0.00897133257240057, + 0.010234609246253967, + 0.011650587432086468, + 0.005275019444525242, + 0.004064627923071384, + 0.008585539646446705, + 0.0055054123513400555, + 0.004920638632029295, + 0.005372973158955574, + 0.00461465073749423, + 0.006057800725102425, + 0.005201576743274927, + 0.010127323679625988, + 0.009046560153365135, + 0.007812864147126675, + 0.015126332640647888, + 0.01988089270889759, + 0.0052734073251485825, + 0.007389694917947054, + 0.009083625860512257, + 0.02718653529882431, + 0.007216235157102346, + 0.009280014783143997, + 0.014975104480981827, + 0.00942239724099636, + 0.0050810277462005615, + 0.005445518530905247, + 0.00444662943482399, + 0.008238589391112328, + 0.006425174418836832, + 0.005595737602561712, + 0.007759771775454283, + 0.01017478946596384, + 0.00480622798204422, + 0.00406618369743228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.0789700523018837 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028152139857411385 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.0028152139857411385, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028152139857411385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06520707160234451, + 0.16270560026168823, + 0.026270732283592224, + 0.04691274091601372, + 0.03713894635438919, + 0.08794484287500381, + 0.03945191577076912, + 0.08470115810632706, + 0.41965728998184204, + 0.03000974841415882 + ], + "credit_map": [ + 0.009991949424147606, + 0.013728223741054535, + 0.0064158313907682896, + 0.007111215963959694, + 0.007189447060227394, + 0.007731618359684944, + 0.006516815163195133, + 0.006521965842694044, + 0.016959046944975853, + 0.02237652614712715, + 0.015306972898542881, + 0.019485965371131897, + 0.027036091312766075, + 0.027225760743021965, + 0.017321180552244186, + 0.016994059085845947, + 0.003770098090171814, + 0.0034729251638054848, + 0.002973238006234169, + 0.002800967311486602, + 0.0037786224856972694, + 0.0037836681585758924, + 0.0030053933151066303, + 0.002685819985345006, + 0.0052785263396799564, + 0.004572833422571421, + 0.008512862958014011, + 0.003640349954366684, + 0.004228776320815086, + 0.007977878674864769, + 0.0041338782757520676, + 0.008567635901272297, + 0.0021260222420096397, + 0.003292754990980029, + 0.004452898167073727, + 0.0050527979619801044, + 0.006591589655727148, + 0.008092246018350124, + 0.004586521536111832, + 0.0029441160149872303, + 0.006269366946071386, + 0.009162704460322857, + 0.0086514288559556, + 0.01025206595659256, + 0.011981531046330929, + 0.011630366556346416, + 0.01335822232067585, + 0.01663915254175663, + 0.004402496386319399, + 0.005513301119208336, + 0.0051516396924853325, + 0.004132852889597416, + 0.005104704760015011, + 0.005399796180427074, + 0.004667594097554684, + 0.005079532973468304, + 0.006746931001543999, + 0.013029344379901886, + 0.008993339724838734, + 0.008138313889503479, + 0.013573657721281052, + 0.016851933673024178, + 0.008790531195700169, + 0.008577102795243263, + 0.019356433302164078, + 0.16576062142848969, + 0.11947713047266006, + 0.019508494064211845, + 0.049856264144182205, + 0.022827502340078354, + 0.009924975223839283, + 0.012945887632668018, + 0.0027495005633682013, + 0.0037200534716248512, + 0.0036825279239565134, + 0.004519334062933922, + 0.00360514665953815, + 0.004100046120584011, + 0.0034016359131783247, + 0.004231502301990986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.0028152139857411385 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed tracking shot, the camera tilting sharply downward, closely follows a pitch-black crow. In an instant, the crow decisively folds its broad wings, tucking them tightly against its sides. Its inky black feathers gleam with a faint metallic sheen in the dim light. Its sturdy beak clenched tightly, its body tensed into a perfectly streamlined form. Then, like a flash of black lightning, it plunges vertically and unhesitatingly toward the blurred ground. A few loose feathers, ripped from its body by the gust of wind, tumble helplessly through the air. Against the backdrop, swirling, leaden storm clouds nearly engulf the sky, only faint, oppressive light glimmering through the gaps in the cloud cover. The entire image is imbued with the grainy, high-contrast quality of vintage film, its tones strictly constrained between varying shades of gray and abject black, creating an atmosphere of tension, coldness, and foreboding doom.", + "seed": 29912, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_1/index_6424_seed_29912_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0436181740835309, + "vlm_logprob": -0.0436181740835309, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.0436181740835309, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12678618729114532 + ], + "opposite_logprob_tokens": [ + -2.251786231994629 + ], + "logprob_reward": -0.12678618729114532, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12678618729114532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0760323777794838, + 0.032952841371297836, + 0.07532650232315063, + 0.07383863627910614, + 0.039135582745075226, + 0.06648661196231842, + 0.03200163319706917, + 0.12205640226602554, + 0.20910361409187317, + 0.27306583523750305 + ], + "credit_map": [ + 0.006995537783950567, + 0.009602962993085384, + 0.0127023383975029, + 0.009173263795673847, + 0.008060022257268429, + 0.010212335735559464, + 0.01163763739168644, + 0.007648278959095478, + 0.005083895288407803, + 0.0031337474938482046, + 0.005825159605592489, + 0.0030378212686628103, + 0.005111757665872574, + 0.003643297590315342, + 0.002845373935997486, + 0.004271788522601128, + 0.014797798357903957, + 0.007052180822938681, + 0.010483663529157639, + 0.008146977052092552, + 0.012218190357089043, + 0.006861353293061256, + 0.008041450753808022, + 0.007724886294454336, + 0.01275070570409298, + 0.007859841920435429, + 0.017189253121614456, + 0.012184876017272472, + 0.004692713730037212, + 0.004830189049243927, + 0.008294939994812012, + 0.0060361223295331, + 0.005934945307672024, + 0.006227537058293819, + 0.007927589118480682, + 0.0042381277307868, + 0.00407382333651185, + 0.004073785152286291, + 0.003099451307207346, + 0.0035603255964815617, + 0.005146707873791456, + 0.015542601235210896, + 0.004511037841439247, + 0.0061399489641189575, + 0.009538930840790272, + 0.017493106424808502, + 0.0038415375165641308, + 0.004272738471627235, + 0.0039835660718381405, + 0.003992239944636822, + 0.003926539793610573, + 0.0038978238590061665, + 0.0030499787535518408, + 0.004563658498227596, + 0.005105824675410986, + 0.0034820009022951126, + 0.020005689933896065, + 0.012775754556059837, + 0.015277436003088951, + 0.014950181357562542, + 0.013241687789559364, + 0.021294614300131798, + 0.014694607816636562, + 0.0098164277151227, + 0.01701449789106846, + 0.02596013993024826, + 0.02173646166920662, + 0.04385145381093025, + 0.020560558885335922, + 0.023066839203238487, + 0.03297165408730507, + 0.023941997438669205, + 0.014253206551074982, + 0.02475903183221817, + 0.039822012186050415, + 0.05791798233985901, + 0.028105255216360092, + 0.023485079407691956, + 0.0644407719373703, + 0.020282484591007233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.12678618729114532 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811003640294075 + ], + "opposite_logprob_tokens": [ + -3.2881100177764893 + ], + "logprob_reward": -0.03811003640294075, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811003640294075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11353839188814163, + 0.1000441312789917, + 0.07399673014879227, + 0.11543712764978409, + 0.049006618559360504, + 0.2573804259300232, + 0.02278456650674343, + 0.0508941188454628, + 0.09552023559808731, + 0.12139769643545151 + ], + "credit_map": [ + 0.014497038908302784, + 0.02230791747570038, + 0.013251492753624916, + 0.01103063765913248, + 0.013156071305274963, + 0.01580309495329857, + 0.011156603693962097, + 0.01233553234487772, + 0.013413000851869583, + 0.016744278371334076, + 0.017341358587145805, + 0.012199713848531246, + 0.008648641407489777, + 0.01492239348590374, + 0.005638411734253168, + 0.011136332526803017, + 0.00953499972820282, + 0.007098618894815445, + 0.007498970255255699, + 0.008087635040283203, + 0.01658160611987114, + 0.008391854353249073, + 0.008175794966518879, + 0.00862724706530571, + 0.022365307435393333, + 0.011561824008822441, + 0.029824266210198402, + 0.014417655766010284, + 0.00977165438234806, + 0.007592742331326008, + 0.008253233507275581, + 0.011650447733700275, + 0.0063066934235394, + 0.0077080377377569675, + 0.010960022918879986, + 0.005032010842114687, + 0.005287235137075186, + 0.004836168140172958, + 0.004331916570663452, + 0.00454452820122242, + 0.014678011648356915, + 0.07285618036985397, + 0.02507118135690689, + 0.07296320796012878, + 0.0192303154617548, + 0.030445335432887077, + 0.012164507992565632, + 0.00997168943285942, + 0.002431799191981554, + 0.00266256439499557, + 0.003530914429575205, + 0.004656706936657429, + 0.0017550346674397588, + 0.0037945148069411516, + 0.0020394613966345787, + 0.0019135716138407588, + 0.00800260715186596, + 0.00497168954461813, + 0.007439343258738518, + 0.005679790861904621, + 0.005784207955002785, + 0.008176027797162533, + 0.00595304649323225, + 0.004887405317276716, + 0.007384144235402346, + 0.012251188978552818, + 0.0092864278703928, + 0.019125038757920265, + 0.009562088176608086, + 0.01266488153487444, + 0.01490872073918581, + 0.010337736457586288, + 0.013962374068796635, + 0.01019398495554924, + 0.0192202627658844, + 0.017169920727610588, + 0.011309772729873657, + 0.011856895871460438, + 0.01754010282456875, + 0.0201443862169981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.03811003640294075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058476023375988 + ], + "opposite_logprob_tokens": [ + -3.89558482170105 + ], + "logprob_reward": -0.02058476023375988, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058476023375988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07742239534854889, + 0.043335821479558945, + 0.11635466665029526, + 0.12091213464736938, + 0.07011602818965912, + 0.06249478831887245, + 0.025679903104901314, + 0.07411820441484451, + 0.1721334606409073, + 0.23743250966072083 + ], + "credit_map": [ + 0.01073553878813982, + 0.014155526645481586, + 0.00916717667132616, + 0.007519911974668503, + 0.009876144118607044, + 0.01001377310603857, + 0.008171443827450275, + 0.0077828881330788136, + 0.005999918561428785, + 0.005826794542372227, + 0.006954800337553024, + 0.003635922446846962, + 0.00639732088893652, + 0.004660259000957012, + 0.003722492605447769, + 0.006138319615274668, + 0.022001666948199272, + 0.013932362198829651, + 0.016747748479247093, + 0.010305135510861874, + 0.02119172364473343, + 0.011207847855985165, + 0.010057562962174416, + 0.010910620912909508, + 0.02198522537946701, + 0.013284971006214619, + 0.03025953844189644, + 0.013536322861909866, + 0.012596799060702324, + 0.007663318887352943, + 0.008097533136606216, + 0.01348844449967146, + 0.011444572359323502, + 0.01081622764468193, + 0.011969992890954018, + 0.0067713698372244835, + 0.00722406804561615, + 0.008759547024965286, + 0.006404973100870848, + 0.006725287064909935, + 0.00595087418332696, + 0.010778476484119892, + 0.006821570917963982, + 0.006850823760032654, + 0.006235620938241482, + 0.016700420528650284, + 0.005448854062706232, + 0.0037081537302583456, + 0.0024361354298889637, + 0.0032478489447385073, + 0.0032246524933725595, + 0.0034360098652541637, + 0.0024443441070616245, + 0.004678217228502035, + 0.002375042997300625, + 0.0038376532029360533, + 0.013148958794772625, + 0.006711276713758707, + 0.008597140200436115, + 0.010042821057140827, + 0.007956269197165966, + 0.008988936431705952, + 0.010717291384935379, + 0.007955520413815975, + 0.012972358614206314, + 0.019706642255187035, + 0.01772359013557434, + 0.03886590152978897, + 0.016063502058386803, + 0.01784493774175644, + 0.029720811173319817, + 0.019235730171203613, + 0.011456617154181004, + 0.02215368114411831, + 0.03714095056056976, + 0.053027257323265076, + 0.025099044665694237, + 0.019126400351524353, + 0.059008482843637466, + 0.010420096106827259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.02058476023375988 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792651534080505 + ], + "opposite_logprob_tokens": [ + -3.529792547225952 + ], + "logprob_reward": -0.029792651534080505, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792651534080505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0510295107960701, + 0.03472360968589783, + 0.04468710720539093, + 0.2615929841995239, + 0.062091898173093796, + 0.07218148559331894, + 0.03674914315342903, + 0.19427211582660675, + 0.15307685732841492, + 0.08959527313709259 + ], + "credit_map": [ + 0.006438331212848425, + 0.0072579653933644295, + 0.008412585593760014, + 0.005851638503372669, + 0.0065874820575118065, + 0.0063171410001814365, + 0.005122130736708641, + 0.005042244680225849, + 0.004778959322720766, + 0.004891594871878624, + 0.004732304252684116, + 0.0039103347808122635, + 0.0037252777256071568, + 0.005003178957849741, + 0.0024952280800789595, + 0.005186730995774269, + 0.007129718083888292, + 0.0053460379131138325, + 0.004576393403112888, + 0.0037919627502560616, + 0.010175250470638275, + 0.004568791016936302, + 0.004098715726286173, + 0.005000241566449404, + 0.04484328255057335, + 0.024908199906349182, + 0.07847340404987335, + 0.03390881046652794, + 0.015395928174257278, + 0.016326120123267174, + 0.023259257897734642, + 0.024477999657392502, + 0.006760932505130768, + 0.010035580955445766, + 0.01509531307965517, + 0.0071012708358466625, + 0.006329658441245556, + 0.006825637072324753, + 0.005610799416899681, + 0.0043327100574970245, + 0.005550360307097435, + 0.013125555589795113, + 0.014697636477649212, + 0.0072698090225458145, + 0.00748424930498004, + 0.015576323494315147, + 0.004311339929699898, + 0.004166219383478165, + 0.003558938857167959, + 0.003922564443200827, + 0.0049985311925411224, + 0.004559253808110952, + 0.003836198477074504, + 0.007401729002594948, + 0.004148184787482023, + 0.004323746543377638, + 0.03133361041545868, + 0.016981102526187897, + 0.03351863473653793, + 0.022298522293567657, + 0.02110590972006321, + 0.02510761469602585, + 0.025450993329286575, + 0.018475746735930443, + 0.012060998007655144, + 0.023042192682623863, + 0.013184446841478348, + 0.028764717280864716, + 0.019118940457701683, + 0.01876593567430973, + 0.02383377030491829, + 0.014305857941508293, + 0.004593080375343561, + 0.009101055562496185, + 0.008730783127248287, + 0.022926049306988716, + 0.008936856873333454, + 0.007138986140489578, + 0.022743327543139458, + 0.005425142124295235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.029792651534080505 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a high-speed tracking shot that tilts sharply downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a pitch-black crow with inky black feathers that have a faint metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow fold its broad wings tightly against its sides and plunge vertically toward the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are loose feathers visible tumbling through the air behind the diving crow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of swirling leaden storm clouds with faint light glimmering through gaps?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817234955728054 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817234955728054, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817234955728054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09662710130214691, + 0.07247751951217651, + 0.053980499505996704, + 0.1618378907442093, + 0.040222521871328354, + 0.1512906849384308, + 0.05425768718123436, + 0.0900389552116394, + 0.12169711291790009, + 0.157570019364357 + ], + "credit_map": [ + 0.011493844911456108, + 0.01665435917675495, + 0.014036351814866066, + 0.010508909821510315, + 0.01423790492117405, + 0.013690472580492496, + 0.008075370453298092, + 0.007929876446723938, + 0.010070453397929668, + 0.010192162357270718, + 0.012082083150744438, + 0.009496698155999184, + 0.006664447020739317, + 0.010147517547011375, + 0.005219766870141029, + 0.008604388684034348, + 0.00849470216780901, + 0.008147790096700191, + 0.007531516719609499, + 0.005785095971077681, + 0.007214602082967758, + 0.006048673298209906, + 0.005127010401338339, + 0.005631110165268183, + 0.02692842110991478, + 0.01728922501206398, + 0.04625070095062256, + 0.023179952055215836, + 0.008747315965592861, + 0.012071633711457253, + 0.012480469420552254, + 0.014890165999531746, + 0.004432366695255041, + 0.007394902873784304, + 0.008288107812404633, + 0.006017626728862524, + 0.003244158113375306, + 0.0051443506963551044, + 0.0025601270608603954, + 0.0031408837530761957, + 0.008513493463397026, + 0.03909855708479881, + 0.018244478851556778, + 0.045041222125291824, + 0.009229774586856365, + 0.01669946499168873, + 0.007844424806535244, + 0.0066192601807415485, + 0.007875327952206135, + 0.00664793374016881, + 0.006174010224640369, + 0.00857497751712799, + 0.004795257933437824, + 0.007920983247458935, + 0.006359254475682974, + 0.005909941159188747, + 0.015599527396261692, + 0.007778596598654985, + 0.013340922072529793, + 0.010705874301493168, + 0.007899921387434006, + 0.015469041652977467, + 0.010027610696852207, + 0.009217462502419949, + 0.006871590856462717, + 0.02940189838409424, + 0.00904625840485096, + 0.01624958962202072, + 0.030527358874678612, + 0.01161200925707817, + 0.010655012913048267, + 0.0073333908803761005, + 0.011769868433475494, + 0.020502451807260513, + 0.03684229403734207, + 0.018720118328928947, + 0.02896087057888508, + 0.016144627705216408, + 0.014192967675626278, + 0.01043681800365448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.002817234955728054 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_10/index_tensor([349, 349])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_10/index_tensor([349, 349])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1bc6ae5bbc5a9dc497f9775a97a2fb6ddc91ca55 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_10/index_tensor([349, 349])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme close-up, the camera remains motionless, its focus fixed squarely on a single, glowing tungsten filament suspended against a profound, pure black background. Heated to incandescence by an electric current, this fine, spiraling metal filament radiates a fiery, brilliant orange-yellow glow, its delicate structure clearly visible in the intense heat. Far from being completely still, it vibrates ever so slightly, constantly, at an almost imperceptible frequency. These minute vibrations imbue the light it emits with a delicate, rhythmic pulsation, like a tiny, quietly breathing organism. This warm glow gently emanates, creating a small, soft-focus halo within the boundless darkness. This striking and captivating contrast creates an atmosphere of focused, microscopic, and almost meditative tranquility, drawing the viewer's attention completely to this tiny, yet inherently powerful source.", + "seed": 27038, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_10/index_349_seed_27038_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.22291567949578167, + "vlm_logprob": -0.22291567949578167, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.22291567949578167, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17397479712963104 + ], + "opposite_logprob_tokens": [ + -2.0489747524261475 + ], + "logprob_reward": -0.17397479712963104, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.17397479712963104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3522408604621887, + 0.10527923703193665, + 0.08743209391832352, + 0.1286032497882843, + 0.061729878187179565, + 0.04212825372815132, + 0.09187808632850647, + 0.007840692065656185, + 0.10961657762527466, + 0.013251073658466339 + ], + "credit_map": [ + 0.03543254733085632, + 0.022461984306573868, + 0.08778475224971771, + 0.04058224335312843, + 0.03473100811243057, + 0.06071186065673828, + 0.036686088889837265, + 0.03385038301348686, + 0.013165711425244808, + 0.00796202290803194, + 0.0185169018805027, + 0.012383957393467426, + 0.011282778345048428, + 0.019713537767529488, + 0.011375842615962029, + 0.01087848749011755, + 0.01316048949956894, + 0.006599355489015579, + 0.015295088291168213, + 0.008022010326385498, + 0.009433235041797161, + 0.01508553046733141, + 0.008657765574753284, + 0.011178618296980858, + 0.019534168764948845, + 0.01138320378959179, + 0.021289564669132233, + 0.011769578792154789, + 0.011984743177890778, + 0.02273966372013092, + 0.013168571516871452, + 0.01673375628888607, + 0.006826638709753752, + 0.0032969429157674313, + 0.012493027374148369, + 0.00958826020359993, + 0.006119341589510441, + 0.01020189467817545, + 0.007056994363665581, + 0.006146779749542475, + 0.0066448082216084, + 0.002983507700264454, + 0.00840199738740921, + 0.003683005692437291, + 0.0031906412914395332, + 0.005801181774586439, + 0.005127717740833759, + 0.0062953936867415905, + 0.009574765339493752, + 0.006009147502481937, + 0.013921065255999565, + 0.012895547784864902, + 0.011225461028516293, + 0.018839595839381218, + 0.010718811303377151, + 0.008693687617778778, + 0.0008340615895576775, + 0.0005007364670746028, + 0.001479790429584682, + 0.00107142876368016, + 0.0009062974131666124, + 0.001541122910566628, + 0.0006953655974939466, + 0.0008118889527395368, + 0.01152702234685421, + 0.010122440755367279, + 0.01660388894379139, + 0.008311895653605461, + 0.013784791342914104, + 0.027103468775749207, + 0.010202338919043541, + 0.011960724368691444, + 0.001446578069590032, + 0.0013296442339196801, + 0.0023455810733139515, + 0.0015333767514675856, + 0.0016825897619128227, + 0.0020427450072020292, + 0.0016619195230305195, + 0.001208639470860362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.17397479712963104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578758001327515 + ], + "opposite_logprob_tokens": [ + -0.6328758001327515 + ], + "logprob_reward": -0.7578758001327515, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578758001327515, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15943528711795807, + 0.035382747650146484, + 0.04062575474381447, + 0.043244969099760056, + 0.019101735204458237, + 0.03359343856573105, + 0.034758277237415314, + 0.13863173127174377, + 0.15735673904418945, + 0.3378692865371704 + ], + "credit_map": [ + 0.015821240842342377, + 0.00984618254005909, + 0.04031900689005852, + 0.017602434381842613, + 0.01336030475795269, + 0.027694277465343475, + 0.01893627643585205, + 0.015855567529797554, + 0.0033138704020529985, + 0.0033483277074992657, + 0.0073239365592598915, + 0.004342592321336269, + 0.0025508578401058912, + 0.0059804487973451614, + 0.004495627712458372, + 0.00402708537876606, + 0.005949895828962326, + 0.0023820006754249334, + 0.006255198735743761, + 0.004670796915888786, + 0.0048798867501318455, + 0.007021138910204172, + 0.00396966515108943, + 0.005497173871845007, + 0.004912906792014837, + 0.0031521234195679426, + 0.008040708489716053, + 0.005085671786218882, + 0.004026039969176054, + 0.007661852054297924, + 0.004306201357394457, + 0.006059464532881975, + 0.002363965380936861, + 0.0014355594757944345, + 0.003208683105185628, + 0.0017860163934528828, + 0.001803465187549591, + 0.0030588635709136724, + 0.003510227194055915, + 0.0019349525682628155, + 0.003610247280448675, + 0.002314183162525296, + 0.00701737217605114, + 0.003205197164788842, + 0.003678929293528199, + 0.0052542476914823055, + 0.0033412768971174955, + 0.005171981640160084, + 0.004106373526155949, + 0.001824027276597917, + 0.007098296657204628, + 0.0038118488155305386, + 0.0032651845831424, + 0.007257347460836172, + 0.004099912941455841, + 0.0032952851615846157, + 0.013050224632024765, + 0.00967391300946474, + 0.01962512917816639, + 0.01596817746758461, + 0.021801073104143143, + 0.02187769114971161, + 0.014925510622560978, + 0.02171001210808754, + 0.014432201161980629, + 0.01531190238893032, + 0.020934894680976868, + 0.017659790813922882, + 0.027989143505692482, + 0.0260566808283329, + 0.019332414492964745, + 0.015639718621969223, + 0.03057049959897995, + 0.032115671783685684, + 0.03965764865279198, + 0.05025649070739746, + 0.05858299508690834, + 0.05396648496389389, + 0.03820258378982544, + 0.034516919404268265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.7578758001327515 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209511682391167 + ], + "opposite_logprob_tokens": [ + -2.8120951652526855 + ], + "logprob_reward": -0.06209511682391167, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209511682391167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16500569880008698, + 0.1687595099210739, + 0.12694773077964783, + 0.0611695796251297, + 0.06275065988302231, + 0.047749172896146774, + 0.09544894099235535, + 0.12279278039932251, + 0.029219938442111015, + 0.12015597522258759 + ], + "credit_map": [ + 0.02586199715733528, + 0.009709415957331657, + 0.033323634415864944, + 0.016050003468990326, + 0.015616931021213531, + 0.03272691369056702, + 0.017858484759926796, + 0.013858318328857422, + 0.029103301465511322, + 0.012175042182207108, + 0.026887957006692886, + 0.021628262475132942, + 0.012803657911717892, + 0.02182033099234104, + 0.024821100756525993, + 0.0195198692381382, + 0.016910165548324585, + 0.006196091417223215, + 0.02059885486960411, + 0.010393012315034866, + 0.011535705998539925, + 0.022653402760624886, + 0.01538230199366808, + 0.023278191685676575, + 0.007257657591253519, + 0.0033984279725700617, + 0.013334778137505054, + 0.006004429422318935, + 0.004450603853911161, + 0.010475787334144115, + 0.007486709859222174, + 0.008761183358728886, + 0.009445782750844955, + 0.005235073156654835, + 0.009360922500491142, + 0.008784712292253971, + 0.006359389051795006, + 0.009977455250918865, + 0.005010982044041157, + 0.008576348423957825, + 0.00454757222905755, + 0.003265018342062831, + 0.009693050757050514, + 0.007904792204499245, + 0.004122898913919926, + 0.007521264720708132, + 0.0053667365573346615, + 0.005327840335667133, + 0.01046051736921072, + 0.007560982834547758, + 0.012295627035200596, + 0.014060157351195812, + 0.012270674109458923, + 0.01329021342098713, + 0.01393758412450552, + 0.011573179624974728, + 0.011992518790066242, + 0.009311978705227375, + 0.01822328194975853, + 0.017714645713567734, + 0.014543724246323109, + 0.02344408631324768, + 0.012917817570269108, + 0.014644728973507881, + 0.0030321443919092417, + 0.0023654995020478964, + 0.00429772911593318, + 0.003922405187040567, + 0.005095839500427246, + 0.004533959086984396, + 0.0035158884711563587, + 0.002456473419442773, + 0.011808328330516815, + 0.009064806625247002, + 0.01635631173849106, + 0.022488674148917198, + 0.01420520804822445, + 0.02124384604394436, + 0.01430287305265665, + 0.010685935616493225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.06209511682391167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300219595432281 + ], + "opposite_logprob_tokens": [ + -2.238002300262451 + ], + "logprob_reward": -0.11300219595432281, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300219595432281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2498035579919815, + 0.11033745110034943, + 0.054818857461214066, + 0.06350170820951462, + 0.11079922318458557, + 0.04701925069093704, + 0.07322058081626892, + 0.06970768421888351, + 0.1179909035563469, + 0.10280079394578934 + ], + "credit_map": [ + 0.026722928509116173, + 0.024228081107139587, + 0.04104869067668915, + 0.031516559422016144, + 0.031482815742492676, + 0.03612503781914711, + 0.032178692519664764, + 0.026500755921006203, + 0.012933062389492989, + 0.009554705582559109, + 0.018837783485651016, + 0.014268826693296432, + 0.01343910675495863, + 0.014856502413749695, + 0.013055991381406784, + 0.013391471467912197, + 0.005270695313811302, + 0.004086821340024471, + 0.00868453923612833, + 0.007223733700811863, + 0.007503087632358074, + 0.00850906316190958, + 0.006566028110682964, + 0.0069748880341649055, + 0.007163587491959333, + 0.004627293441444635, + 0.014828008599579334, + 0.008141636848449707, + 0.005629475694149733, + 0.009789415635168552, + 0.00706610269844532, + 0.00625618314370513, + 0.014293348416686058, + 0.00693971524015069, + 0.018675627186894417, + 0.011698075570166111, + 0.011885055340826511, + 0.01739174872636795, + 0.014153185300529003, + 0.015762463212013245, + 0.008009428158402443, + 0.0029784704092890024, + 0.009320035576820374, + 0.004839679691940546, + 0.0033830420579761267, + 0.006178621668368578, + 0.00487296050414443, + 0.0074370140209794044, + 0.009162175469100475, + 0.0040718624368309975, + 0.011221378110349178, + 0.00973806343972683, + 0.007645351346582174, + 0.01443553902208805, + 0.007808472495526075, + 0.009137733839452267, + 0.008610064163804054, + 0.004234418272972107, + 0.010356955230236053, + 0.007258345372974873, + 0.008597034029662609, + 0.010414904914796352, + 0.008133264258503914, + 0.012102703563869, + 0.011086118407547474, + 0.010689642280340195, + 0.0162167027592659, + 0.013670156709849834, + 0.017921030521392822, + 0.023380395025014877, + 0.013446170836687088, + 0.0115806944668293, + 0.008076772093772888, + 0.009236834943294525, + 0.01422248687595129, + 0.01610470749437809, + 0.015930160880088806, + 0.016969209536910057, + 0.012340179644525051, + 0.009920453652739525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.11300219595432281 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630487438291311 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630487438291311, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630487438291311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09122731536626816, + 0.05170629546046257, + 0.12287965416908264, + 0.12313347309827805, + 0.06907616555690765, + 0.037853509187698364, + 0.03708560764789581, + 0.06079718470573425, + 0.049180835485458374, + 0.35705992579460144 + ], + "credit_map": [ + 0.011933413334190845, + 0.008297220803797245, + 0.016907231882214546, + 0.008163891732692719, + 0.007093349937349558, + 0.01471265684813261, + 0.012330278754234314, + 0.011789282783865929, + 0.0066787186078727245, + 0.0040973578579723835, + 0.00767964543774724, + 0.007822902873158455, + 0.005395408719778061, + 0.007595623377710581, + 0.006448125466704369, + 0.005988513585180044, + 0.02217485010623932, + 0.007818812504410744, + 0.02321782335639, + 0.009401489049196243, + 0.01089506596326828, + 0.0208344217389822, + 0.012982405722141266, + 0.015554787591099739, + 0.01501101441681385, + 0.007981270551681519, + 0.02385777421295643, + 0.015859730541706085, + 0.010522067546844482, + 0.019829709082841873, + 0.012739044614136219, + 0.017332859337329865, + 0.0079493448138237, + 0.005728726275265217, + 0.010413517244160175, + 0.01114741712808609, + 0.006300956942141056, + 0.011402602307498455, + 0.0076154181733727455, + 0.00851818360388279, + 0.006362643558532, + 0.002087696921080351, + 0.007267302833497524, + 0.003624899312853813, + 0.003032851731404662, + 0.0052495370618999004, + 0.004548833705484867, + 0.00567974429577589, + 0.004802769050002098, + 0.0023938899394124746, + 0.005569566506892443, + 0.004332270007580519, + 0.004673860035836697, + 0.006449317093938589, + 0.004880465101450682, + 0.003983470611274242, + 0.006045269314199686, + 0.004317950922995806, + 0.010219336487352848, + 0.01097057107836008, + 0.007463310845196247, + 0.008775325492024422, + 0.006662268657237291, + 0.006343154236674309, + 0.005309421569108963, + 0.004491847939789295, + 0.006707918830215931, + 0.0069843498058617115, + 0.007658143062144518, + 0.007501407526433468, + 0.005894605536013842, + 0.0046331374906003475, + 0.03687655180692673, + 0.027606487274169922, + 0.04516839608550072, + 0.08630236983299255, + 0.03614955022931099, + 0.044409286230802536, + 0.049512941390275955, + 0.03103436343371868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.007630487438291311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera remains motionless, its focus fixed squarely on a single, glowing tungsten filament suspended against a profound, pure black background. Heated to incandescence by an electric current, this fine, spiraling metal filament radiates a fiery, brilliant orange-yellow glow, its delicate structure clearly visible in the intense heat. Far from being completely still, it vibrates ever so slightly, constantly, at an almost imperceptible frequency. These minute vibrations imbue the light it emits with a delicate, rhythmic pulsation, like a tiny, quietly breathing organism. This warm glow gently emanates, creating a small, soft-focus halo within the boundless darkness. This striking and captivating contrast creates an atmosphere of focused, microscopic, and almost meditative tranquility, drawing the viewer's attention completely to this tiny, yet inherently powerful source.", + "seed": 27039, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_10/index_349_seed_27039_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1287761738989502, + "vlm_logprob": -0.1287761738989502, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.1287761738989502, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12865473330020905 + ], + "opposite_logprob_tokens": [ + -2.378654718399048 + ], + "logprob_reward": -0.12865473330020905, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12865473330020905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.238610178232193, + 0.07643883675336838, + 0.031680695712566376, + 0.08040300756692886, + 0.07815206050872803, + 0.08976341784000397, + 0.09788932651281357, + 0.13139010965824127, + 0.15616537630558014, + 0.019507024437189102 + ], + "credit_map": [ + 0.02161732316017151, + 0.02243545651435852, + 0.04756224900484085, + 0.027291269972920418, + 0.021409664303064346, + 0.0374712273478508, + 0.03920392692089081, + 0.02161906659603119, + 0.007870491594076157, + 0.007283009588718414, + 0.012040221132338047, + 0.010698790661990643, + 0.008298197761178017, + 0.008320351131260395, + 0.014410978183150291, + 0.007516791578382254, + 0.0034861573949456215, + 0.0034618726931512356, + 0.005239052698016167, + 0.0032126486767083406, + 0.0035928625147789717, + 0.003398865694180131, + 0.005742756184190512, + 0.0035464786924421787, + 0.011084145866334438, + 0.008736527524888515, + 0.013170365244150162, + 0.008647496812045574, + 0.006271334830671549, + 0.012111905962228775, + 0.01244916208088398, + 0.007932066917419434, + 0.012290993705391884, + 0.009687297977507114, + 0.014301777817308903, + 0.0084993252530694, + 0.0060066767036914825, + 0.008990105241537094, + 0.01001513097435236, + 0.008360755629837513, + 0.018125716596841812, + 0.011268483474850655, + 0.01795145310461521, + 0.009412430226802826, + 0.004321620799601078, + 0.009696259163320065, + 0.00932819303125143, + 0.009659266099333763, + 0.013103543780744076, + 0.01057723630219698, + 0.019001739099621773, + 0.011364229023456573, + 0.00562300905585289, + 0.012159178964793682, + 0.01675598882138729, + 0.009304402396082878, + 0.01138348039239645, + 0.013311201706528664, + 0.02485000342130661, + 0.011301984079182148, + 0.011453812010586262, + 0.0241865124553442, + 0.021017076447606087, + 0.013886034488677979, + 0.019222630187869072, + 0.01418531034141779, + 0.028998831287026405, + 0.013822871260344982, + 0.009677903726696968, + 0.02754531428217888, + 0.024909228086471558, + 0.017803288996219635, + 0.002049289410933852, + 0.0021906658075749874, + 0.0037034612614661455, + 0.0022307252511382103, + 0.0018535215640440583, + 0.002142199082300067, + 0.003330162027850747, + 0.0020069994498044252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.12865473330020905 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289577007293701 + ], + "opposite_logprob_tokens": [ + -1.0539577007293701 + ], + "logprob_reward": -0.4289577007293701, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289577007293701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06191369891166687, + 0.057969532907009125, + 0.061153825372457504, + 0.15489187836647034, + 0.3658011555671692, + 0.19678542017936707, + 0.022332454100251198, + 0.02342916838824749, + 0.018731212243437767, + 0.03699163720011711 + ], + "credit_map": [ + 0.007810095325112343, + 0.004249169025570154, + 0.014132112264633179, + 0.0050836605951189995, + 0.004874737001955509, + 0.009247975423932076, + 0.009467760100960732, + 0.0070481873117387295, + 0.00573910353705287, + 0.00529780238866806, + 0.010511336848139763, + 0.006271108984947205, + 0.005578816402703524, + 0.00687987869605422, + 0.011167124845087528, + 0.006524365395307541, + 0.007261678110808134, + 0.008175029419362545, + 0.012411682866513729, + 0.006948668044060469, + 0.005745831876993179, + 0.008348910138010979, + 0.007645703852176666, + 0.0046163215301930904, + 0.013661687262356281, + 0.010386364534497261, + 0.031299930065870285, + 0.012673250399529934, + 0.013892237097024918, + 0.03156863525509834, + 0.02096875011920929, + 0.020441018044948578, + 0.031584762036800385, + 0.031513139605522156, + 0.08065441995859146, + 0.03033412992954254, + 0.02749556303024292, + 0.06727644056081772, + 0.055769115686416626, + 0.041173577308654785, + 0.017450271174311638, + 0.02653639391064644, + 0.03948180004954338, + 0.019856050610542297, + 0.01511690765619278, + 0.03216250240802765, + 0.026187613606452942, + 0.01999387890100479, + 0.0023773869033902884, + 0.0030995351262390614, + 0.0029174501542001963, + 0.0028796130791306496, + 0.002214506734162569, + 0.002515834290534258, + 0.004453019239008427, + 0.0018751071766018867, + 0.002783376956358552, + 0.0022111404687166214, + 0.0049266270361840725, + 0.0020871025044471025, + 0.0017388993874192238, + 0.00446760468184948, + 0.003041772171854973, + 0.0021726449485868216, + 0.0022305650636553764, + 0.0019808700308203697, + 0.003743232460692525, + 0.002122259698808193, + 0.0017484532436355948, + 0.0021790815517306328, + 0.002870701253414154, + 0.0018560487078502774, + 0.003457944840192795, + 0.0033675371669232845, + 0.006070590578019619, + 0.003992334008216858, + 0.0033193924464285374, + 0.00449099438264966, + 0.00827433355152607, + 0.0040185097604990005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.4289577007293701 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592818036675453 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592818036675453, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592818036675453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20216818153858185, + 0.10722841322422028, + 0.03573824092745781, + 0.08039312064647675, + 0.09904978424310684, + 0.0903363898396492, + 0.14585842192173004, + 0.11208425462245941, + 0.07625072449445724, + 0.05089247226715088 + ], + "credit_map": [ + 0.037003032863140106, + 0.015563727356493473, + 0.0350366048514843, + 0.02080833725631237, + 0.011445815674960613, + 0.024357395246624947, + 0.026266038417816162, + 0.03168724104762077, + 0.017344065010547638, + 0.01142111886292696, + 0.024735860526561737, + 0.00895711500197649, + 0.008011849597096443, + 0.01726766675710678, + 0.010381657630205154, + 0.009109072387218475, + 0.0034129666164517403, + 0.003981661517173052, + 0.005800997838377953, + 0.003380695590749383, + 0.003207308240234852, + 0.0046022431924939156, + 0.005990267731249332, + 0.005362102296203375, + 0.013679185882210732, + 0.0075212144292891026, + 0.01367019023746252, + 0.00793208833783865, + 0.006638044025748968, + 0.00881833303719759, + 0.012564249336719513, + 0.009569813497364521, + 0.014433071948587894, + 0.009891345165669918, + 0.021375253796577454, + 0.01036500558257103, + 0.005863884463906288, + 0.017063913866877556, + 0.00932554341852665, + 0.010731762275099754, + 0.010695381090044975, + 0.012153368443250656, + 0.015183472074568272, + 0.011262818239629269, + 0.008343304507434368, + 0.009194946847856045, + 0.015417465940117836, + 0.00808563083410263, + 0.017147710546851158, + 0.020479371771216393, + 0.027599897235631943, + 0.01813524402678013, + 0.01368172001093626, + 0.014331125654280186, + 0.02125782147049904, + 0.013225517235696316, + 0.012373857200145721, + 0.011377857066690922, + 0.018636003136634827, + 0.012316846288740635, + 0.011530384421348572, + 0.012262317351996899, + 0.020765550434589386, + 0.0128214405849576, + 0.008439682424068451, + 0.008372554555535316, + 0.014927276410162449, + 0.007623708341270685, + 0.005060996860265732, + 0.01382903940975666, + 0.010349910706281662, + 0.007647558115422726, + 0.009332084096968174, + 0.004950311034917831, + 0.010194229893386364, + 0.005135219078511, + 0.0035838123876601458, + 0.0062667387537658215, + 0.0059853908605873585, + 0.005444681737571955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.020592818036675453 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206105649471283 + ], + "opposite_logprob_tokens": [ + -2.812061071395874 + ], + "logprob_reward": -0.06206105649471283, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206105649471283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4804740250110626, + 0.11233334243297577, + 0.05542735755443573, + 0.04818184673786163, + 0.05910190939903259, + 0.03025176376104355, + 0.04760083556175232, + 0.08955393731594086, + 0.047085728496313095, + 0.029989248141646385 + ], + "credit_map": [ + 0.04131422936916351, + 0.04400613158941269, + 0.09614349901676178, + 0.055440325289964676, + 0.04722825810313225, + 0.07349567860364914, + 0.08438193053007126, + 0.03846396878361702, + 0.012432835064828396, + 0.010470626875758171, + 0.017858874052762985, + 0.01545313373208046, + 0.012206049636006355, + 0.010550440289080143, + 0.023537447676062584, + 0.009823940694332123, + 0.006406902335584164, + 0.005961827002465725, + 0.00865422934293747, + 0.006943908985704184, + 0.005386861972510815, + 0.005722610279917717, + 0.011708115227520466, + 0.00464290427044034, + 0.0036854275967925787, + 0.003320164978504181, + 0.010154728777706623, + 0.005624242592602968, + 0.0037637092173099518, + 0.00785136315971613, + 0.008517866022884846, + 0.005264343693852425, + 0.00515717500820756, + 0.005659554619342089, + 0.010405008681118488, + 0.005722506903111935, + 0.004896538332104683, + 0.00974556989967823, + 0.011703941971063614, + 0.005811616778373718, + 0.002317467937245965, + 0.0026678545400500298, + 0.006443370599299669, + 0.0037684838753193617, + 0.0023322165943682194, + 0.003984855487942696, + 0.006028368137776852, + 0.002709146123379469, + 0.005290352739393711, + 0.0062550390139222145, + 0.008437804877758026, + 0.005321239586919546, + 0.003701605834066868, + 0.006219685077667236, + 0.007486520800739527, + 0.004888589959591627, + 0.008611470460891724, + 0.010394626297056675, + 0.017162702977657318, + 0.010289453901350498, + 0.008184097707271576, + 0.013107642531394958, + 0.01593664288520813, + 0.005867303349077702, + 0.0048400056548416615, + 0.004085211548954248, + 0.008569283410906792, + 0.004587458446621895, + 0.003343879710882902, + 0.00739840604364872, + 0.008939540944993496, + 0.005321937147527933, + 0.0041613271459937096, + 0.0025874257553368807, + 0.0054341391660273075, + 0.002832029014825821, + 0.0020228831563144922, + 0.004536780063062906, + 0.005266961641609669, + 0.0031477005686610937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.06206105649471283 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036145609337836504 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036145609337836504, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036145609337836504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28298288583755493, + 0.0945764034986496, + 0.09979979693889618, + 0.06641814112663269, + 0.07327491790056229, + 0.08412312716245651, + 0.06725378334522247, + 0.11273372918367386, + 0.09095419198274612, + 0.027882982045412064 + ], + "credit_map": [ + 0.026419861242175102, + 0.02792944759130478, + 0.054100558161735535, + 0.028591178357601166, + 0.030216220766305923, + 0.04075353592634201, + 0.04896519333124161, + 0.02600691094994545, + 0.009732239879667759, + 0.008678417652845383, + 0.016799967736005783, + 0.01257137581706047, + 0.008883437141776085, + 0.010187550447881222, + 0.017973611131310463, + 0.009749799966812134, + 0.0076723964884877205, + 0.01090184785425663, + 0.022597188130021095, + 0.010846875607967377, + 0.011544526554644108, + 0.012805592268705368, + 0.015409818850457668, + 0.008021545596420765, + 0.007559215184301138, + 0.008228049613535404, + 0.014134427532553673, + 0.00698066595941782, + 0.0073946015909314156, + 0.006968583911657333, + 0.010250942781567574, + 0.004901653155684471, + 0.0054102251306176186, + 0.006470395717769861, + 0.019102441146969795, + 0.008016210049390793, + 0.0057668318040668964, + 0.011872638016939163, + 0.010314002633094788, + 0.00632217712700367, + 0.008649650029838085, + 0.01240853127092123, + 0.01375649031251669, + 0.010747777298092842, + 0.00598551332950592, + 0.011633781716227531, + 0.014732671901583672, + 0.0062087043188512325, + 0.0072008115239441395, + 0.010371692478656769, + 0.014569656923413277, + 0.0075210644863545895, + 0.0052865720354020596, + 0.009855685755610466, + 0.007036670111119747, + 0.0054116263054311275, + 0.007871094159781933, + 0.013220757246017456, + 0.01969161070883274, + 0.014756128191947937, + 0.010459307581186295, + 0.012888389639556408, + 0.024182014167308807, + 0.00966443307697773, + 0.00839499942958355, + 0.00884732510894537, + 0.015758000314235687, + 0.008786345832049847, + 0.006758070085197687, + 0.015203646384179592, + 0.01631953939795494, + 0.01088627614080906, + 0.0034178243950009346, + 0.0027720662765204906, + 0.004732172004878521, + 0.002874762751162052, + 0.0017671455862000585, + 0.004229542799293995, + 0.005069173872470856, + 0.0030202940106391907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0036145609337836504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera remains motionless, its focus fixed squarely on a single, glowing tungsten filament suspended against a profound, pure black background. Heated to incandescence by an electric current, this fine, spiraling metal filament radiates a fiery, brilliant orange-yellow glow, its delicate structure clearly visible in the intense heat. Far from being completely still, it vibrates ever so slightly, constantly, at an almost imperceptible frequency. These minute vibrations imbue the light it emits with a delicate, rhythmic pulsation, like a tiny, quietly breathing organism. This warm glow gently emanates, creating a small, soft-focus halo within the boundless darkness. This striking and captivating contrast creates an atmosphere of focused, microscopic, and almost meditative tranquility, drawing the viewer's attention completely to this tiny, yet inherently powerful source.", + "seed": 27040, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_10/index_349_seed_27040_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1647225096821785, + "vlm_logprob": -0.1647225096821785, + "vlm_yes_no_margin": 3.424999952316284, + "vlm_token_credit": -0.1647225096821785, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029221227392554283 + ], + "opposite_logprob_tokens": [ + -4.029221057891846 + ], + "logprob_reward": -0.029221227392554283, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.029221227392554283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2419705092906952, + 0.06930401921272278, + 0.05599144846200943, + 0.03581678867340088, + 0.04354923591017723, + 0.0706740990281105, + 0.14947251975536346, + 0.0831446722149849, + 0.12699514627456665, + 0.12308157980442047 + ], + "credit_map": [ + 0.03490034118294716, + 0.03072523884475231, + 0.04456154257059097, + 0.03906436264514923, + 0.0164492204785347, + 0.020900823175907135, + 0.02065086178481579, + 0.034718107432127, + 0.010698288679122925, + 0.008719749748706818, + 0.011716870591044426, + 0.011021963320672512, + 0.005477251019328833, + 0.005416864063590765, + 0.007217308972030878, + 0.00903572142124176, + 0.007354618050158024, + 0.005971482954919338, + 0.01000180747359991, + 0.010143867693841457, + 0.0029204702004790306, + 0.003949117846786976, + 0.0037429367657750845, + 0.01190714817494154, + 0.003973609767854214, + 0.004401481710374355, + 0.006643418688327074, + 0.007332746870815754, + 0.0022135283797979355, + 0.0027490081265568733, + 0.003055284498259425, + 0.00544770946726203, + 0.005893254652619362, + 0.005343260243535042, + 0.008724710904061794, + 0.006626315414905548, + 0.0030748280696570873, + 0.0036523640155792236, + 0.0040800077840685844, + 0.006154494825750589, + 0.008335870690643787, + 0.011965462937951088, + 0.009453642182052135, + 0.020460428670048714, + 0.0026526562869548798, + 0.004643373657017946, + 0.003543739439919591, + 0.009618928655982018, + 0.016925636678934097, + 0.023509055376052856, + 0.026280807331204414, + 0.03466014191508293, + 0.005692151375114918, + 0.009356684051454067, + 0.0064529296942055225, + 0.026595113798975945, + 0.012945679947733879, + 0.011065898463129997, + 0.015171715058386326, + 0.01327900867909193, + 0.005854964721947908, + 0.007393745705485344, + 0.006106081418693066, + 0.011327574960887432, + 0.01727580651640892, + 0.017189033329486847, + 0.029727593064308167, + 0.02195795439183712, + 0.00818744394928217, + 0.008875563740730286, + 0.011720333248376846, + 0.01206141896545887, + 0.015751754865050316, + 0.01946326531469822, + 0.017915816977620125, + 0.02428182028234005, + 0.007521279156208038, + 0.009427961893379688, + 0.006734169088304043, + 0.021985506638884544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.029221227392554283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578460574150085 + ], + "opposite_logprob_tokens": [ + -0.6328460574150085 + ], + "logprob_reward": -0.7578460574150085, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578460574150085, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17548108100891113, + 0.06118714064359665, + 0.049719084054231644, + 0.06316056847572327, + 0.025378452613949776, + 0.042993735522031784, + 0.06563396006822586, + 0.17904624342918396, + 0.15170560777187347, + 0.18569405376911163 + ], + "credit_map": [ + 0.026353778317570686, + 0.020700521767139435, + 0.030881617218255997, + 0.029860123991966248, + 0.008332915604114532, + 0.011539514176547527, + 0.0129685765132308, + 0.034844037145376205, + 0.008711264468729496, + 0.00751636503264308, + 0.008004890754818916, + 0.009813908487558365, + 0.005858955439180136, + 0.006644520442932844, + 0.005574900656938553, + 0.00906234048306942, + 0.006685270927846432, + 0.007159132044762373, + 0.008079081773757935, + 0.009337563067674637, + 0.004398764576762915, + 0.0038321788888424635, + 0.0042563872411847115, + 0.005970704834908247, + 0.006949263624846935, + 0.00729228975251317, + 0.012652729637920856, + 0.012744012288749218, + 0.004069903865456581, + 0.0043550943955779076, + 0.006866473238915205, + 0.00823080725967884, + 0.0033475470263510942, + 0.003534184768795967, + 0.004477818962186575, + 0.0036859428510069847, + 0.0015161834890022874, + 0.0020782246720045805, + 0.0020594927482306957, + 0.004679056815803051, + 0.004778976086527109, + 0.006172059569507837, + 0.006747911684215069, + 0.008081239648163319, + 0.0035990094766020775, + 0.004149516578763723, + 0.0039020904805511236, + 0.005562932696193457, + 0.007447321433573961, + 0.00898012425750494, + 0.013116382993757725, + 0.011180820874869823, + 0.004273497499525547, + 0.005659007467329502, + 0.005445942282676697, + 0.009530861862003803, + 0.02890327200293541, + 0.016873493790626526, + 0.029657065868377686, + 0.02934689074754715, + 0.015766490250825882, + 0.016540560871362686, + 0.016939643770456314, + 0.02501882240176201, + 0.016442924737930298, + 0.01818843185901642, + 0.034800682216882706, + 0.024673696607351303, + 0.010550404898822308, + 0.01198434829711914, + 0.012664190493524075, + 0.022400932386517525, + 0.02079685591161251, + 0.022585123777389526, + 0.032031573355197906, + 0.044989246875047684, + 0.010713070631027222, + 0.014629453420639038, + 0.010151151567697525, + 0.029797593131661415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.7578460574150085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421281136572361 + ], + "opposite_logprob_tokens": [ + -4.264212608337402 + ], + "logprob_reward": -0.01421281136572361, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421281136572361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38247355818748474, + 0.0644763708114624, + 0.07052356004714966, + 0.08239269256591797, + 0.07237475365400314, + 0.16760322451591492, + 0.022423507645726204, + 0.0692891851067543, + 0.0202146228402853, + 0.04822850599884987 + ], + "credit_map": [ + 0.05512077733874321, + 0.049296606332063675, + 0.0629776269197464, + 0.06165049597620964, + 0.02534283883869648, + 0.03766496106982231, + 0.03640614077448845, + 0.054014112800359726, + 0.011463995091617107, + 0.008330359123647213, + 0.00991404615342617, + 0.00929378904402256, + 0.0034761670976877213, + 0.00549611821770668, + 0.0049230619333684444, + 0.011578832753002644, + 0.010588349774479866, + 0.010110833682119846, + 0.011098652146756649, + 0.01277061365544796, + 0.005380378570407629, + 0.00598145741969347, + 0.005950937047600746, + 0.008642340078949928, + 0.010610893368721008, + 0.012679986655712128, + 0.013320929370820522, + 0.019325362518429756, + 0.004726443439722061, + 0.005926195997744799, + 0.0062604499980807304, + 0.009542432613670826, + 0.012951746582984924, + 0.00862673670053482, + 0.01352037861943245, + 0.010344174690544605, + 0.0037772604264318943, + 0.006575493607670069, + 0.005395249463617802, + 0.01118371170014143, + 0.020879214629530907, + 0.024985257536172867, + 0.025572888553142548, + 0.040339864790439606, + 0.00753713259473443, + 0.013472549617290497, + 0.008471429347991943, + 0.026344886049628258, + 0.0038047712296247482, + 0.0023915215861052275, + 0.0033050114288926125, + 0.003252715803682804, + 0.0014327445533126593, + 0.00236411951482296, + 0.00234354380518198, + 0.0035290801897644997, + 0.010633555240929127, + 0.00823717750608921, + 0.013653910718858242, + 0.011367498897016048, + 0.004619997460395098, + 0.006337778177112341, + 0.00511843292042613, + 0.009320840239524841, + 0.002399853430688381, + 0.002440096577629447, + 0.0036797714419662952, + 0.002747825114056468, + 0.0015773858176544309, + 0.002326252404600382, + 0.0017424063989892602, + 0.003301032353192568, + 0.0055291373282670975, + 0.007796634919941425, + 0.007527707144618034, + 0.010474201291799545, + 0.0030215277802199125, + 0.0038501268718391657, + 0.003466782160103321, + 0.006562387105077505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.01421281136572361 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979368481785059 + ], + "opposite_logprob_tokens": [ + -4.634793758392334 + ], + "logprob_reward": -0.00979368481785059, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979368481785059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09930098056793213, + 0.11152331531047821, + 0.07628285139799118, + 0.08289824426174164, + 0.10434485971927643, + 0.1643403321504593, + 0.06058451160788536, + 0.10371117293834686, + 0.12963415682315826, + 0.06737960129976273 + ], + "credit_map": [ + 0.012334535829722881, + 0.011236804537475109, + 0.015595107339322567, + 0.01693282276391983, + 0.006809105630964041, + 0.008627696894109249, + 0.009405178017914295, + 0.018359718844294548, + 0.014215804636478424, + 0.012093105353415012, + 0.015447827987372875, + 0.01933133229613304, + 0.011042867787182331, + 0.011706612072885036, + 0.011294269002974033, + 0.016391484066843987, + 0.009884798899292946, + 0.010361270979046822, + 0.012089699506759644, + 0.0147625133395195, + 0.0059822313487529755, + 0.006860633846372366, + 0.00796530582010746, + 0.008376392535865307, + 0.01243501529097557, + 0.010893157683312893, + 0.013784470036625862, + 0.016351548954844475, + 0.005943146999925375, + 0.007853365503251553, + 0.007745629642158747, + 0.007891899906098843, + 0.01196278352290392, + 0.019043339416384697, + 0.0178986769169569, + 0.029253065586090088, + 0.003760045161470771, + 0.005890074651688337, + 0.004478873685002327, + 0.01205800473690033, + 0.02610931731760502, + 0.02213294804096222, + 0.02417139895260334, + 0.03572151064872742, + 0.0073209903202950954, + 0.013183949515223503, + 0.012061028741300106, + 0.023639174178242683, + 0.006543068680912256, + 0.0063602253794670105, + 0.007791820913553238, + 0.010318935848772526, + 0.00497937947511673, + 0.005719337612390518, + 0.0054456377401947975, + 0.013426100835204124, + 0.011503135785460472, + 0.012527037411928177, + 0.018681218847632408, + 0.017800720408558846, + 0.00927120540291071, + 0.008552432991564274, + 0.01136446837335825, + 0.014010944403707981, + 0.014073476195335388, + 0.019177453592419624, + 0.024317128583788872, + 0.02652190625667572, + 0.010118802078068256, + 0.010931240394711494, + 0.010985284112393856, + 0.013508862815797329, + 0.009324473328888416, + 0.006624126341193914, + 0.01212263572961092, + 0.008801511488854885, + 0.005116126500070095, + 0.005515759345144033, + 0.005236092954874039, + 0.014638873748481274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.00979368481785059 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538767419755459 + ], + "opposite_logprob_tokens": [ + -4.387538909912109 + ], + "logprob_reward": -0.012538767419755459, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538767419755459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15848909318447113, + 0.07221827656030655, + 0.11333256214857101, + 0.06282112002372742, + 0.23183707892894745, + 0.1645958572626114, + 0.034760359674692154, + 0.05130188167095184, + 0.031637001782655716, + 0.07900678366422653 + ], + "credit_map": [ + 0.02760804072022438, + 0.0160740427672863, + 0.02334795519709587, + 0.022153107449412346, + 0.006322501227259636, + 0.011240215972065926, + 0.009767591953277588, + 0.04197563976049423, + 0.00860094465315342, + 0.009392564184963703, + 0.011181790381669998, + 0.015844635665416718, + 0.003983963746577501, + 0.005531188566237688, + 0.0038817760068923235, + 0.013801410794258118, + 0.017301859334111214, + 0.014445185661315918, + 0.01867436058819294, + 0.022327831014990807, + 0.004695631098002195, + 0.007758963853120804, + 0.0052008372731506824, + 0.022927895188331604, + 0.011079157702624798, + 0.008935109712183475, + 0.013201911933720112, + 0.011015303432941437, + 0.003229451132938266, + 0.0038772851694375277, + 0.003033621236681938, + 0.008449279703199863, + 0.02806040272116661, + 0.03676044940948486, + 0.032893259078264236, + 0.07615498453378677, + 0.007188329938799143, + 0.013041479513049126, + 0.013918587937951088, + 0.023819582536816597, + 0.02076335810124874, + 0.02709944359958172, + 0.02410600520670414, + 0.04853981360793114, + 0.0048085711896419525, + 0.008891823701560497, + 0.009959439747035503, + 0.020427407696843147, + 0.004860514309257269, + 0.005290822591632605, + 0.005913335829973221, + 0.0085179153829813, + 0.0016017608577385545, + 0.0020263746846467257, + 0.002427315805107355, + 0.00412231869995594, + 0.007664832286536694, + 0.006996127776801586, + 0.009921442717313766, + 0.009825512766838074, + 0.0031695212237536907, + 0.0047387839294970036, + 0.003636052133515477, + 0.005349607206881046, + 0.0043283021077513695, + 0.0036835428327322006, + 0.006303785368800163, + 0.004167203325778246, + 0.002520900219678879, + 0.002545731607824564, + 0.002911168849095702, + 0.005176363978534937, + 0.008922455832362175, + 0.01077937800437212, + 0.013220918364822865, + 0.019507965072989464, + 0.002473923610523343, + 0.004973134025931358, + 0.0032318488229066133, + 0.01589716412127018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.012538767419755459 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera remains motionless, its focus fixed squarely on a single, glowing tungsten filament suspended against a profound, pure black background. Heated to incandescence by an electric current, this fine, spiraling metal filament radiates a fiery, brilliant orange-yellow glow, its delicate structure clearly visible in the intense heat. Far from being completely still, it vibrates ever so slightly, constantly, at an almost imperceptible frequency. These minute vibrations imbue the light it emits with a delicate, rhythmic pulsation, like a tiny, quietly breathing organism. This warm glow gently emanates, creating a small, soft-focus halo within the boundless darkness. This striking and captivating contrast creates an atmosphere of focused, microscopic, and almost meditative tranquility, drawing the viewer's attention completely to this tiny, yet inherently powerful source.", + "seed": 27041, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_10/index_349_seed_27041_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16020910302177072, + "vlm_logprob": -0.16020910302177072, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.16020910302177072, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019280405715107918 + ], + "opposite_logprob_tokens": [ + -5.019280433654785 + ], + "logprob_reward": -0.019280405715107918, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.019280405715107918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5169879198074341, + 0.11943458765745163, + 0.05355251580476761, + 0.06847747415304184, + 0.03186911717057228, + 0.024554818868637085, + 0.061766721308231354, + 0.05573170259594917, + 0.03514494001865387, + 0.03248019516468048 + ], + "credit_map": [ + 0.035393718630075455, + 0.018421141430735588, + 0.15028950572013855, + 0.03506937250494957, + 0.032298970967531204, + 0.15948468446731567, + 0.024362051859498024, + 0.06166850030422211, + 0.011632331646978855, + 0.007800170686095953, + 0.024385148659348488, + 0.011481350287795067, + 0.01714274100959301, + 0.01923402212560177, + 0.009145031683146954, + 0.0186137892305851, + 0.005988255608826876, + 0.0032678667921572924, + 0.009009026922285557, + 0.004713632632046938, + 0.006060099229216576, + 0.008318646810948849, + 0.004017713479697704, + 0.012177271768450737, + 0.008934696204960346, + 0.005710157100111246, + 0.015572845935821533, + 0.006205987185239792, + 0.006076687015593052, + 0.009964071214199066, + 0.0057093179784715176, + 0.010303712449967861, + 0.004257241729646921, + 0.002701676683500409, + 0.0052270847372710705, + 0.0028746756725013256, + 0.00333870155736804, + 0.004413440357893705, + 0.0025186080019921064, + 0.006537687499076128, + 0.00321508408524096, + 0.0023085547145456076, + 0.005060251802206039, + 0.00254565617069602, + 0.0028770866338163614, + 0.0038643067236989737, + 0.0018230032874271274, + 0.0028608760330826044, + 0.007731960155069828, + 0.00300780963152647, + 0.01448800414800644, + 0.0039271521382033825, + 0.004422641359269619, + 0.010934177786111832, + 0.003667469136416912, + 0.013587506487965584, + 0.0074464245699346066, + 0.0038419216871261597, + 0.009044085629284382, + 0.004181661643087864, + 0.00633613346144557, + 0.0076799094676971436, + 0.003883017459884286, + 0.01331854797899723, + 0.0034196970518678427, + 0.0035444488748908043, + 0.00600627064704895, + 0.0033303925301879644, + 0.005751532036811113, + 0.004595916252583265, + 0.0033851279877126217, + 0.005111555103212595, + 0.0036245151422917843, + 0.002889875788241625, + 0.007310620509088039, + 0.003467392874881625, + 0.00364156742580235, + 0.005171490833163261, + 0.002780906856060028, + 0.0035938238725066185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.019280405715107918 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934012770652771 + ], + "opposite_logprob_tokens": [ + -0.6934012770652771 + ], + "logprob_reward": -0.6934012770652771, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934012770652771, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2130138874053955, + 0.0688166543841362, + 0.06678038090467453, + 0.07469114661216736, + 0.21429583430290222, + 0.18748678267002106, + 0.06301425397396088, + 0.035114407539367676, + 0.03660997003316879, + 0.04017673805356026 + ], + "credit_map": [ + 0.018799448385834694, + 0.02170158177614212, + 0.03351235389709473, + 0.032407622784376144, + 0.025525404140353203, + 0.03086266852915287, + 0.019939156249165535, + 0.03026563674211502, + 0.0062829540111124516, + 0.008224353194236755, + 0.015232201665639877, + 0.008991635404527187, + 0.006076492834836245, + 0.008003256283700466, + 0.0059232343919575214, + 0.010082532651722431, + 0.006944867316633463, + 0.007634757086634636, + 0.012331684119999409, + 0.007245727814733982, + 0.00864407792687416, + 0.007526500150561333, + 0.005957300774753094, + 0.010495469905436039, + 0.008227544836699963, + 0.00515978317707777, + 0.015034974552690983, + 0.007417484652251005, + 0.011359015479683876, + 0.010982146486639977, + 0.007217722944915295, + 0.00929247122257948, + 0.022440290078520775, + 0.01615639217197895, + 0.037929363548755646, + 0.030664747580885887, + 0.028869008645415306, + 0.02835747040808201, + 0.01907501183450222, + 0.030803553760051727, + 0.022568093612790108, + 0.018242405727505684, + 0.028737390413880348, + 0.025768883526325226, + 0.023313144221901894, + 0.022404801100492477, + 0.014332546852529049, + 0.0321195162832737, + 0.010463119484484196, + 0.0034417861606925726, + 0.010037701576948166, + 0.004332329612225294, + 0.003677438013255596, + 0.008068361319601536, + 0.0034069204702973366, + 0.0195865947753191, + 0.00341817201115191, + 0.004506933037191629, + 0.005015082191675901, + 0.00409682746976614, + 0.004173319321125746, + 0.00430394196882844, + 0.0035877209156751633, + 0.006012411322444677, + 0.0036397743970155716, + 0.0032920637167990208, + 0.007078548427671194, + 0.004218485672026873, + 0.004545194562524557, + 0.006173057481646538, + 0.002868067706003785, + 0.00479477783665061, + 0.00364498607814312, + 0.004990280140191317, + 0.006720472127199173, + 0.004912152886390686, + 0.006785392761230469, + 0.004841447342187166, + 0.004771946929395199, + 0.003510056994855404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.6934012770652771 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006305456161499 + ], + "opposite_logprob_tokens": [ + -2.6950631141662598 + ], + "logprob_reward": -0.07006305456161499, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006305456161499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14554442465305328, + 0.11361654847860336, + 0.1585494428873062, + 0.04840000718832016, + 0.08392194658517838, + 0.06633605808019638, + 0.1814342588186264, + 0.13025598227977753, + 0.02503715455532074, + 0.04690421372652054 + ], + "credit_map": [ + 0.018994377925992012, + 0.010246312245726585, + 0.0301250908523798, + 0.01284941378980875, + 0.01576857455074787, + 0.017410796135663986, + 0.012492900714278221, + 0.027656950056552887, + 0.013491133227944374, + 0.005035639740526676, + 0.03192102536559105, + 0.007800766732543707, + 0.006559591740369797, + 0.01954646222293377, + 0.005764658562839031, + 0.023497266694903374, + 0.023292135447263718, + 0.011919158510863781, + 0.044558364897966385, + 0.012239503674209118, + 0.00878740381449461, + 0.029713407158851624, + 0.009380836971104145, + 0.01865861378610134, + 0.005829125177115202, + 0.002876274986192584, + 0.00793797243386507, + 0.005054486449807882, + 0.003623589174821973, + 0.006151433568447828, + 0.0037106070667505264, + 0.013216514140367508, + 0.011071339249610901, + 0.005922647193074226, + 0.014613335020840168, + 0.006538873538374901, + 0.008623444475233555, + 0.011127211153507233, + 0.005944033153355122, + 0.020081058144569397, + 0.006748799234628677, + 0.005202716216444969, + 0.013181411661207676, + 0.007129755336791277, + 0.010481243021786213, + 0.009684043005108833, + 0.005157848820090294, + 0.008750233799219131, + 0.01915779337286949, + 0.01564943790435791, + 0.031533945351839066, + 0.014245289377868176, + 0.03725295886397362, + 0.025368565693497658, + 0.01541552972048521, + 0.02281073108315468, + 0.012671212665736675, + 0.011649076826870441, + 0.020410114899277687, + 0.010877669788897038, + 0.0277106910943985, + 0.01781182922422886, + 0.011622386984527111, + 0.017503002658486366, + 0.0019477035384625196, + 0.002710169181227684, + 0.005139578599482775, + 0.0026952289044857025, + 0.003749817842617631, + 0.003837228287011385, + 0.0022255650255829096, + 0.0027318596839904785, + 0.0071432627737522125, + 0.002894166624173522, + 0.008742593228816986, + 0.0034027332440018654, + 0.003497160505503416, + 0.006218368653208017, + 0.0032178249675780535, + 0.011788100004196167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.07006305456161499 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206582680344582 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206582680344582, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206582680344582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15382172167301178, + 0.16829867660999298, + 0.24341227114200592, + 0.14738070964813232, + 0.059451524168252945, + 0.06531199812889099, + 0.04398318752646446, + 0.02757333405315876, + 0.052586186677217484, + 0.03818037733435631 + ], + "credit_map": [ + 0.013964169658720493, + 0.017511477693915367, + 0.022640695795416832, + 0.015113313682377338, + 0.023152871057391167, + 0.02157563529908657, + 0.01618783362209797, + 0.023675713688135147, + 0.017123771831393242, + 0.014107553288340569, + 0.037450145930051804, + 0.016531402245163918, + 0.023289211094379425, + 0.027520842850208282, + 0.013531135395169258, + 0.01874460279941559, + 0.02508005127310753, + 0.017032591626048088, + 0.06437230855226517, + 0.021272700279951096, + 0.03307099640369415, + 0.04434219375252724, + 0.01623191311955452, + 0.022009504958987236, + 0.015394223853945732, + 0.01117468811571598, + 0.03177616000175476, + 0.013718575239181519, + 0.019080935046076775, + 0.025026176124811172, + 0.010186595842242241, + 0.021023346111178398, + 0.005365839693695307, + 0.006095235701650381, + 0.00985138863325119, + 0.008126037195324898, + 0.008084449917078018, + 0.007868709973990917, + 0.005130291450768709, + 0.008929569274187088, + 0.006797606125473976, + 0.00507782306522131, + 0.01159739214926958, + 0.006387130822986364, + 0.006861818954348564, + 0.010136098600924015, + 0.004048994276672602, + 0.014405133202672005, + 0.006925388239324093, + 0.002946964930742979, + 0.007055746391415596, + 0.003280516481027007, + 0.004276416264474392, + 0.006181004457175732, + 0.00235345889814198, + 0.010963689535856247, + 0.0024892687797546387, + 0.002648360561579466, + 0.0045922864228487015, + 0.0029945827554911375, + 0.0032605419401079416, + 0.003802312072366476, + 0.002189462771639228, + 0.005596513859927654, + 0.0059203095734119415, + 0.003625387093052268, + 0.011457953602075577, + 0.0043064262717962265, + 0.005123875569552183, + 0.008881943300366402, + 0.0034423505421727896, + 0.009827937930822372, + 0.0031549183186143637, + 0.0029800920747220516, + 0.008120502345263958, + 0.002772497246041894, + 0.004448612220585346, + 0.005390651058405638, + 0.0026190089993178844, + 0.008694094605743885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.014206582680344582 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the filament exhibit a fine, spiraling structure that vibrates slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the light emitted create a small, soft-focus halo in the darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094195086508989 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094195086508989, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094195086508989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1971120983362198, + 0.12960295379161835, + 0.1471536010503769, + 0.05362522974610329, + 0.10826942324638367, + 0.031077295541763306, + 0.1141723096370697, + 0.04220276698470116, + 0.04699739068746567, + 0.1297869086265564 + ], + "credit_map": [ + 0.018502946943044662, + 0.01249206718057394, + 0.05035778135061264, + 0.01845334656536579, + 0.026851991191506386, + 0.037884362041950226, + 0.01645834371447563, + 0.016111254692077637, + 0.014683516696095467, + 0.012193797156214714, + 0.02680855430662632, + 0.01328146643936634, + 0.017120853066444397, + 0.02290818840265274, + 0.010015122592449188, + 0.012591451406478882, + 0.018111685290932655, + 0.016637850552797318, + 0.03267979621887207, + 0.015126222744584084, + 0.013636723160743713, + 0.020310552790760994, + 0.013751310296356678, + 0.016899460926651955, + 0.006705648265779018, + 0.005768883507698774, + 0.009443719871342182, + 0.0038870631251484156, + 0.00611941097304225, + 0.0074896439909935, + 0.004750611260533333, + 0.009460249915719032, + 0.016193199902772903, + 0.010548759251832962, + 0.01982153207063675, + 0.011467956006526947, + 0.009176732040941715, + 0.014157123863697052, + 0.00972831156104803, + 0.017175814136862755, + 0.004412197507917881, + 0.003156813094392419, + 0.005887162871658802, + 0.0034069593530148268, + 0.003899375209584832, + 0.0037883473560214043, + 0.0022133304737508297, + 0.004313110373914242, + 0.012166558764874935, + 0.009753271006047726, + 0.021562904119491577, + 0.009015954099595547, + 0.022338081151247025, + 0.017530737444758415, + 0.009746870025992393, + 0.012057928368449211, + 0.004609405528753996, + 0.0034827357158064842, + 0.005821638274937868, + 0.004227291792631149, + 0.00765339145436883, + 0.006090170703828335, + 0.004211840685456991, + 0.006106293294578791, + 0.00649740407243371, + 0.0025341215077787638, + 0.009932315908372402, + 0.002552150748670101, + 0.003971501719206572, + 0.006656021345406771, + 0.0026598141994327307, + 0.01219406258314848, + 0.015009772032499313, + 0.010105635039508343, + 0.021861715242266655, + 0.014403209090232849, + 0.009752985090017319, + 0.01608862355351448, + 0.012029873207211494, + 0.030535085126757622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.004094195086508989 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_11/index_tensor([1206, 1206])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_11/index_tensor([1206, 1206])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ff042182c1c12d3ef7010b3f1fcf7a1642c93d8e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_11/index_tensor([1206, 1206])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static close-up depicts a woman on the left side of the screen, her presence imbuing the image with warmth and expressiveness. Her honey-blonde curls fall over the shoulders of a soft, cream-colored cashmere sweater. Warm, natural light from an off-screen source illuminates the right side of her face, highlighting her fair complexion and kind brown eyes. The background to her right is softly out of focus, revealing the blurred shapes of a bookshelf and the gentle glow of a warm, small lamp, creating a cozy and intimate interior. As she speaks, her expression is vivid and captivating; she occasionally blinks slowly, nods thoughtfully, and her gaze briefly shifts to the right of the screen, as if recalling a detail. For emphasis, she deliberately raises her eyebrows, then lets them settle. She then leans forward slightly, a genuine, warm smile spreading across her face as she continues her point, drawing the viewer closer to the conversation. Finally, she straightens her posture, her body relaxing, and her head gently sways from side to side in a comfortable rhythm, conveying a genuine and joyful atmosphere of conversation.", + "seed": 4626, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_11/index_1206_seed_4626_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.054813084006309507, + "vlm_logprob": -0.054813084006309507, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.054813084006309507, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13108471035957336 + ], + "opposite_logprob_tokens": [ + -2.131084680557251 + ], + "logprob_reward": -0.13108471035957336, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13108471035957336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17588578164577484, + 0.06603344529867172, + 0.09102925658226013, + 0.04322538524866104, + 0.0872371569275856, + 0.13523036241531372, + 0.09459242969751358, + 0.08284478634595871, + 0.16361276805400848, + 0.06030863896012306 + ], + "credit_map": [ + 0.018940022215247154, + 0.028068097308278084, + 0.01756347343325615, + 0.022366173565387726, + 0.026685651391744614, + 0.014184574596583843, + 0.02783418633043766, + 0.02024359628558159, + 0.005677231587469578, + 0.0077152918092906475, + 0.0071420688182115555, + 0.008883041329681873, + 0.014186171814799309, + 0.005598687566816807, + 0.00989475753158331, + 0.006936191115528345, + 0.011924857273697853, + 0.005818333011120558, + 0.009251951240003109, + 0.015063726343214512, + 0.009935715235769749, + 0.00658252602443099, + 0.025265458971261978, + 0.007186689414083958, + 0.005666838958859444, + 0.005051527172327042, + 0.004349768161773682, + 0.006798789836466312, + 0.007437457330524921, + 0.003941148519515991, + 0.004813936073333025, + 0.005165919661521912, + 0.006249092984944582, + 0.011755338869988918, + 0.00868296343833208, + 0.016376037150621414, + 0.011966627091169357, + 0.005718739703297615, + 0.01815606839954853, + 0.008332290686666965, + 0.008074031211435795, + 0.017024492844939232, + 0.01221489254385233, + 0.015359379351139069, + 0.026742277666926384, + 0.011937758885324001, + 0.015495672821998596, + 0.028381846845149994, + 0.007986831478774548, + 0.009395867586135864, + 0.020837385207414627, + 0.013554660603404045, + 0.015441091731190681, + 0.009671961888670921, + 0.006818376947194338, + 0.010886250995099545, + 0.0036853894125670195, + 0.006253731902688742, + 0.006426024716347456, + 0.009855235926806927, + 0.02052505686879158, + 0.020469030365347862, + 0.004130476154386997, + 0.011499845422804356, + 0.017170248553156853, + 0.020905131474137306, + 0.01635480858385563, + 0.03398846462368965, + 0.022667618468403816, + 0.016817452386021614, + 0.01745106652379036, + 0.018257973715662956, + 0.004209533333778381, + 0.009472050704061985, + 0.004020842723548412, + 0.011752164922654629, + 0.010203824378550053, + 0.011863470077514648, + 0.005599742289632559, + 0.0031870133243501186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.13108471035957336 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026411330327391624 + ], + "opposite_logprob_tokens": [ + -3.651411294937134 + ], + "logprob_reward": -0.026411330327391624, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026411330327391624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1332850456237793, + 0.07997056096792221, + 0.030511198565363884, + 0.020874392241239548, + 0.10686860233545303, + 0.08064234256744385, + 0.05958190932869911, + 0.06258322298526764, + 0.08543418347835541, + 0.34024858474731445 + ], + "credit_map": [ + 0.01521200593560934, + 0.009733120910823345, + 0.01019984483718872, + 0.020169813185930252, + 0.017454251646995544, + 0.009697692468762398, + 0.035069309175014496, + 0.0157490037381649, + 0.0123392129316926, + 0.007664748467504978, + 0.009947599843144417, + 0.0095681082457304, + 0.006808171514421701, + 0.0053862896747887135, + 0.017086906358599663, + 0.01116951834410429, + 0.0063416981138288975, + 0.0029942700639367104, + 0.0023083952255547047, + 0.003886064048856497, + 0.0031842170283198357, + 0.0026400245260447264, + 0.0047766538336873055, + 0.004379874560981989, + 0.0019911490380764008, + 0.002632712246850133, + 0.0032371566630899906, + 0.002995522925630212, + 0.0035072683822363615, + 0.0014193807728588581, + 0.0026307932566851377, + 0.00246041058562696, + 0.008665726520121098, + 0.020309150218963623, + 0.006793905049562454, + 0.01658743806183338, + 0.014984493143856525, + 0.009512907825410366, + 0.017402445897459984, + 0.012612533755600452, + 0.004186115227639675, + 0.013356880284845829, + 0.009210359305143356, + 0.009978209622204304, + 0.014000004157423973, + 0.007065406534820795, + 0.009444805793464184, + 0.013400562107563019, + 0.005467372015118599, + 0.006946309003978968, + 0.007444067858159542, + 0.009485590271651745, + 0.009816262871026993, + 0.004965794738382101, + 0.005899607203900814, + 0.00955690536648035, + 0.002826901851221919, + 0.004841035231947899, + 0.0062544443644583225, + 0.006853814236819744, + 0.016420096158981323, + 0.015626993030309677, + 0.0036251831334084272, + 0.006134761963039637, + 0.009588397108018398, + 0.010096020065248013, + 0.008671914227306843, + 0.01998923160135746, + 0.013288966380059719, + 0.006889170967042446, + 0.009550269693136215, + 0.00736021576449275, + 0.0175632331520319, + 0.054097749292850494, + 0.021010562777519226, + 0.04180002212524414, + 0.06488563120365143, + 0.06026430428028107, + 0.04488836228847504, + 0.03573872148990631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.026411330327391624 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867260530591011 + ], + "opposite_logprob_tokens": [ + -3.048672676086426 + ], + "logprob_reward": -0.04867260530591011, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867260530591011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08014126867055893, + 0.24217185378074646, + 0.08514303714036942, + 0.039745181798934937, + 0.3345205783843994, + 0.061856459826231, + 0.04184262827038765, + 0.01890913024544716, + 0.04532776027917862, + 0.050342146307229996 + ], + "credit_map": [ + 0.01274741068482399, + 0.007223953492939472, + 0.008970395661890507, + 0.011203006841242313, + 0.010570296086370945, + 0.0061979894526302814, + 0.01633371412754059, + 0.0068944962695240974, + 0.03806187957525253, + 0.02299620769917965, + 0.029070036485791206, + 0.028909912332892418, + 0.02293374389410019, + 0.016157442703843117, + 0.04994907230138779, + 0.03409356623888016, + 0.015743138268589973, + 0.006906805094331503, + 0.008367116563022137, + 0.013669776730239391, + 0.007437687832862139, + 0.006975973024964333, + 0.014504166319966316, + 0.011538378894329071, + 0.005677695386111736, + 0.004184576217085123, + 0.005006995052099228, + 0.005955847445875406, + 0.006000397261232138, + 0.0034166336990892887, + 0.005436476785689592, + 0.004066562745720148, + 0.02403245121240616, + 0.0632450059056282, + 0.02203056402504444, + 0.0577697791159153, + 0.043630003929138184, + 0.028662974014878273, + 0.0644683763384819, + 0.030681412667036057, + 0.003796627977862954, + 0.011312232352793217, + 0.00505604175850749, + 0.007481611333787441, + 0.010978021658957005, + 0.0055353376083076, + 0.005565647501498461, + 0.012130938470363617, + 0.004735252819955349, + 0.004993626847863197, + 0.0046908059157431126, + 0.005932573229074478, + 0.005993591155856848, + 0.003515802090987563, + 0.004468260332942009, + 0.0075127179734408855, + 0.0015424203593283892, + 0.002929995534941554, + 0.002385272178798914, + 0.004363995511084795, + 0.0025016036815941334, + 0.0014154700329527259, + 0.001170444069430232, + 0.0025999315548688173, + 0.0029619988054037094, + 0.004106073174625635, + 0.004961208440363407, + 0.008294638246297836, + 0.006225853692740202, + 0.005722380243241787, + 0.006920155603438616, + 0.0061354488134384155, + 0.004508344456553459, + 0.008918577805161476, + 0.0039061373099684715, + 0.008531030267477036, + 0.008919203653931618, + 0.004900802858173847, + 0.006642632186412811, + 0.004015419166535139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.04867260530591011 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097988814115524 + ], + "opposite_logprob_tokens": [ + -3.288098096847534 + ], + "logprob_reward": -0.038097988814115524, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097988814115524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07313437759876251, + 0.07621175795793533, + 0.02926689200103283, + 0.026860980316996574, + 0.09555960446596146, + 0.12591254711151123, + 0.0378151535987854, + 0.0602743960916996, + 0.12503458559513092, + 0.34992972016334534 + ], + "credit_map": [ + 0.00755425076931715, + 0.005936597939580679, + 0.00907424371689558, + 0.010992839932441711, + 0.011055517010390759, + 0.005754351150244474, + 0.01580963097512722, + 0.006956943776458502, + 0.009406231343746185, + 0.007743166293948889, + 0.006677310913801193, + 0.010306701064109802, + 0.007951032370328903, + 0.005861183628439903, + 0.01753351092338562, + 0.010732613503932953, + 0.0024056658148765564, + 0.002865969203412533, + 0.0031324075534939766, + 0.005336238536983728, + 0.0036019065883010626, + 0.002084125066176057, + 0.006677735131233931, + 0.0031628417782485485, + 0.002979880664497614, + 0.0032554329372942448, + 0.0024315910413861275, + 0.005078807473182678, + 0.004136660136282444, + 0.0022242579143494368, + 0.003927987068891525, + 0.0028263612184673548, + 0.008102103136479855, + 0.019242146983742714, + 0.010216115042567253, + 0.018133623525500298, + 0.011448832228779793, + 0.007339330855756998, + 0.012723436579108238, + 0.008354014717042446, + 0.010311095975339413, + 0.026542922481894493, + 0.012713042087852955, + 0.028612127527594566, + 0.016737019643187523, + 0.008260007947683334, + 0.008809427730739117, + 0.013926904648542404, + 0.004090355709195137, + 0.005171388387680054, + 0.006601619999855757, + 0.005636157933622599, + 0.0057400912046432495, + 0.003836818505078554, + 0.002808730583637953, + 0.003929989878088236, + 0.003723980626091361, + 0.005413834471255541, + 0.006749317515641451, + 0.009318962693214417, + 0.011874860152602196, + 0.012205366976559162, + 0.0034962936770170927, + 0.007491779513657093, + 0.015839923173189163, + 0.016758600249886513, + 0.020750636234879494, + 0.02333424612879753, + 0.016169046983122826, + 0.010061554610729218, + 0.013470364734530449, + 0.008650196716189384, + 0.026984916999936104, + 0.061384085565805435, + 0.027168704196810722, + 0.08554831147193909, + 0.048753149807453156, + 0.03925928846001625, + 0.037757810205221176, + 0.02307344786822796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.038097988814115524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798785224556923 + ], + "opposite_logprob_tokens": [ + -3.529798746109009 + ], + "logprob_reward": -0.029798785224556923, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798785224556923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10213558375835419, + 0.08592898398637772, + 0.05093274638056755, + 0.03991641476750374, + 0.18242214620113373, + 0.0358307845890522, + 0.05234207585453987, + 0.0913257971405983, + 0.2101181298494339, + 0.1490473449230194 + ], + "credit_map": [ + 0.01944419927895069, + 0.008564387448132038, + 0.0102073447778821, + 0.013835488818585873, + 0.012131045572459698, + 0.006758510135114193, + 0.020191000774502754, + 0.011003602296113968, + 0.005798198748379946, + 0.008738089352846146, + 0.006392011418938637, + 0.013771521858870983, + 0.011421766132116318, + 0.006676717195659876, + 0.02347608283162117, + 0.00965458806604147, + 0.004296311642974615, + 0.0040965331718325615, + 0.00496693653985858, + 0.009493441320955753, + 0.005771593190729618, + 0.0036447669845074415, + 0.01370266079902649, + 0.00496050575748086, + 0.005975397769361734, + 0.004626917187124491, + 0.004446417558938265, + 0.005085551179945469, + 0.0055566453374922276, + 0.0033381753601133823, + 0.0047778477892279625, + 0.006109464447945356, + 0.012568429112434387, + 0.020051412284374237, + 0.018861964344978333, + 0.03738902136683464, + 0.021473251283168793, + 0.01255184318870306, + 0.044141486287117004, + 0.015384752303361893, + 0.0017506720032542944, + 0.0053977565839886665, + 0.0035932862665504217, + 0.009188743308186531, + 0.005658630281686783, + 0.0030094238463789225, + 0.0031752402428537607, + 0.004057032056152821, + 0.004944512154906988, + 0.006546718068420887, + 0.011294781230390072, + 0.010046614333987236, + 0.008013984188437462, + 0.0036348856519907713, + 0.004443716257810593, + 0.0034168658312410116, + 0.004182965029031038, + 0.008758442476391792, + 0.009854213334619999, + 0.025876231491565704, + 0.017308037728071213, + 0.013130463659763336, + 0.005185852292925119, + 0.0070295920595526695, + 0.02487114444375038, + 0.023685771971940994, + 0.03132433071732521, + 0.047988418489694595, + 0.0286126509308815, + 0.021344957873225212, + 0.021360302343964577, + 0.010930554009974003, + 0.013875357806682587, + 0.028093857690691948, + 0.013479587621986866, + 0.0402878075838089, + 0.02017270028591156, + 0.010153461247682571, + 0.012816437520086765, + 0.01016814075410366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.029798785224556923 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up depicts a woman on the left side of the screen, her presence imbuing the image with warmth and expressiveness. Her honey-blonde curls fall over the shoulders of a soft, cream-colored cashmere sweater. Warm, natural light from an off-screen source illuminates the right side of her face, highlighting her fair complexion and kind brown eyes. The background to her right is softly out of focus, revealing the blurred shapes of a bookshelf and the gentle glow of a warm, small lamp, creating a cozy and intimate interior. As she speaks, her expression is vivid and captivating; she occasionally blinks slowly, nods thoughtfully, and her gaze briefly shifts to the right of the screen, as if recalling a detail. For emphasis, she deliberately raises her eyebrows, then lets them settle. She then leans forward slightly, a genuine, warm smile spreading across her face as she continues her point, drawing the viewer closer to the conversation. Finally, she straightens her posture, her body relaxing, and her head gently sways from side to side in a comfortable rhythm, conveying a genuine and joyful atmosphere of conversation.", + "seed": 4627, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_11/index_1206_seed_4627_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07665350739844143, + "vlm_logprob": -0.07665350739844143, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.07665350739844143, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003053647233173251 + ], + "opposite_logprob_tokens": [ + -6.628053665161133 + ], + "logprob_reward": -0.003053647233173251, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.003053647233173251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42939355969429016, + 0.04413437098264694, + 0.05274690315127373, + 0.051253076642751694, + 0.04024377465248108, + 0.03849336504936218, + 0.044824082404375076, + 0.04918113723397255, + 0.06467703729867935, + 0.18505273759365082 + ], + "credit_map": [ + 0.06480969488620758, + 0.04607723280787468, + 0.06719211488962173, + 0.05778186768293381, + 0.02465776354074478, + 0.059486255049705505, + 0.043205853551626205, + 0.06618277728557587, + 0.007072464097291231, + 0.004849425051361322, + 0.00448076194152236, + 0.005293610505759716, + 0.004771315958350897, + 0.005620703101158142, + 0.005601606331765652, + 0.006444480735808611, + 0.00762183777987957, + 0.00364693533629179, + 0.005636563990265131, + 0.00628703786060214, + 0.003066357923671603, + 0.012637014500796795, + 0.004144406411796808, + 0.009706750512123108, + 0.00473543256521225, + 0.005025721155107021, + 0.007469597272574902, + 0.00533282058313489, + 0.004220637492835522, + 0.009135201573371887, + 0.005224417895078659, + 0.010109247639775276, + 0.007098819594830275, + 0.004676965065300465, + 0.005690332036465406, + 0.0030624072533100843, + 0.002900929655879736, + 0.005472507793456316, + 0.004813574720174074, + 0.006528238300234079, + 0.0051363809034228325, + 0.004815832711756229, + 0.004181312397122383, + 0.004878539592027664, + 0.0037506495136767626, + 0.005016807466745377, + 0.005081352777779102, + 0.005632487591356039, + 0.006540804635733366, + 0.006474792491644621, + 0.0036001568660140038, + 0.006655633449554443, + 0.00426346343010664, + 0.005936192348599434, + 0.00453418493270874, + 0.006818857509642839, + 0.008600876666605473, + 0.007042481563985348, + 0.004706589505076408, + 0.00784844160079956, + 0.0033610162790864706, + 0.0053641474805772305, + 0.005157835781574249, + 0.007099745329469442, + 0.011051198467612267, + 0.007104127202183008, + 0.008781744167208672, + 0.008029909804463387, + 0.0056956312619149685, + 0.008330626413226128, + 0.007567014079540968, + 0.008116785436868668, + 0.015296555124223232, + 0.016643362119793892, + 0.01482369564473629, + 0.01384415477514267, + 0.019718946889042854, + 0.05405102297663689, + 0.016111470758914948, + 0.03456353023648262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.003053647233173251 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016100041568279266 + ], + "opposite_logprob_tokens": [ + -4.14109992980957 + ], + "logprob_reward": -0.016100041568279266, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016100041568279266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23486191034317017, + 0.06747565418481827, + 0.09084974229335785, + 0.0870252400636673, + 0.1273137480020523, + 0.06942778080701828, + 0.052221182733774185, + 0.0754326656460762, + 0.11783494800329208, + 0.07755714654922485 + ], + "credit_map": [ + 0.022051341831684113, + 0.026650168001651764, + 0.04180973395705223, + 0.03568915277719498, + 0.017668621614575386, + 0.03340061381459236, + 0.0220126211643219, + 0.035579659044742584, + 0.008629560470581055, + 0.007910133339464664, + 0.010586810298264027, + 0.01125914603471756, + 0.005036826711148024, + 0.00786808505654335, + 0.007174461614340544, + 0.00901062972843647, + 0.013010441325604916, + 0.007599819917231798, + 0.010324846021831036, + 0.009868964552879333, + 0.007071562577039003, + 0.018562961369752884, + 0.011400296352803707, + 0.013010854832828045, + 0.009970608167350292, + 0.006922271568328142, + 0.006603826303035021, + 0.006063275970518589, + 0.005900640040636063, + 0.022483661770820618, + 0.0087854890152812, + 0.020295463502407074, + 0.016170309856534004, + 0.009867426939308643, + 0.011720765382051468, + 0.011926958337426186, + 0.012992793694138527, + 0.023084010928869247, + 0.014008485712110996, + 0.027542993426322937, + 0.010673034004867077, + 0.005372128915041685, + 0.006454887334257364, + 0.00493607809767127, + 0.0042564598843455315, + 0.017126912251114845, + 0.008146400563418865, + 0.0124618886038661, + 0.006487733218818903, + 0.006007818039506674, + 0.007805691100656986, + 0.008846302516758442, + 0.006201313808560371, + 0.006291496567428112, + 0.005551772657781839, + 0.005029055755585432, + 0.00810061115771532, + 0.006721358746290207, + 0.007351060397922993, + 0.008990840055048466, + 0.006112226285040379, + 0.018011685460805893, + 0.008304214105010033, + 0.01184067316353321, + 0.01683664508163929, + 0.012253284454345703, + 0.015381503850221634, + 0.010577731765806675, + 0.013318220153450966, + 0.023220207542181015, + 0.010688734240829945, + 0.015558627434074879, + 0.009840328246355057, + 0.010362869128584862, + 0.008935471996665001, + 0.009809569455683231, + 0.005605059675872326, + 0.016778476536273956, + 0.007309092208743095, + 0.008916281163692474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.016100041568279266 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631197106093168 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631197106093168, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631197106093168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33173033595085144, + 0.11586347967386246, + 0.11224047094583511, + 0.02021322213113308, + 0.01738499477505684, + 0.050714679062366486, + 0.040223900228738785, + 0.08243343979120255, + 0.07800878584384918, + 0.1511867344379425 + ], + "credit_map": [ + 0.05308151617646217, + 0.03323117643594742, + 0.03202465921640396, + 0.04375831410288811, + 0.04647698625922203, + 0.049459490925073624, + 0.03438669443130493, + 0.039311483502388, + 0.012620650231838226, + 0.009512703865766525, + 0.00918829906731844, + 0.011636562645435333, + 0.019557712599635124, + 0.02071344293653965, + 0.014139736071228981, + 0.018494371324777603, + 0.013590087182819843, + 0.006962434854358435, + 0.008088200353085995, + 0.009607461281120777, + 0.02180011011660099, + 0.024586142972111702, + 0.010078092105686665, + 0.017527949064970016, + 0.0027887816540896893, + 0.002072759438306093, + 0.0023388792760670185, + 0.002184410812333226, + 0.0027935125399380922, + 0.0024634255096316338, + 0.0027262696530669928, + 0.002845184179022908, + 0.002728146966546774, + 0.0023510020691901445, + 0.0020779164042323828, + 0.0017388276755809784, + 0.0012613438302651048, + 0.002949455752968788, + 0.0017879502847790718, + 0.002490353537723422, + 0.006051954347640276, + 0.0031360993161797523, + 0.0036239009350538254, + 0.0026170294731855392, + 0.007523436099290848, + 0.012166166678071022, + 0.0057679410092532635, + 0.009828152135014534, + 0.0045341020449995995, + 0.005161641165614128, + 0.003831123234704137, + 0.003939628601074219, + 0.004232392646372318, + 0.006311515346169472, + 0.0039518228732049465, + 0.008261675015091896, + 0.009465605951845646, + 0.008291377685964108, + 0.008810164406895638, + 0.010021024383604527, + 0.007057931274175644, + 0.019669394940137863, + 0.008150630630552769, + 0.010967308655381203, + 0.011590584181249142, + 0.011717468500137329, + 0.010372624732553959, + 0.009625673294067383, + 0.005342194810509682, + 0.008139731362462044, + 0.008966334164142609, + 0.012254168279469013, + 0.015253785997629166, + 0.025201287120580673, + 0.017699670046567917, + 0.014477924443781376, + 0.02010437659919262, + 0.020412113517522812, + 0.019628142938017845, + 0.01840943656861782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.007631197106093168 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134171962738037 + ], + "opposite_logprob_tokens": [ + -1.3134171962738037 + ], + "logprob_reward": -0.3134171962738037, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134171962738037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09210135042667389, + 0.059852465987205505, + 0.035137154161930084, + 0.0461634024977684, + 0.032082609832286835, + 0.02870592474937439, + 0.056731995195150375, + 0.14182880520820618, + 0.24386219680309296, + 0.2635341286659241 + ], + "credit_map": [ + 0.011319379322230816, + 0.010732636786997318, + 0.01142215821892023, + 0.009944412857294083, + 0.007625556085258722, + 0.01591448113322258, + 0.00933864712715149, + 0.015804078429937363, + 0.008220464922487736, + 0.0067377169616520405, + 0.005564344115555286, + 0.00956894364207983, + 0.004343150649219751, + 0.008463178761303425, + 0.005958363879472017, + 0.010996298864483833, + 0.0037584395613521338, + 0.003583380253985524, + 0.003341477829962969, + 0.0037183412350714207, + 0.0027180016040802, + 0.008626746013760567, + 0.0028825642075389624, + 0.006508202757686377, + 0.008535845205187798, + 0.004247449804097414, + 0.007012213114649057, + 0.004543281625956297, + 0.002656613476574421, + 0.006788268219679594, + 0.0036975634284317493, + 0.00868216622620821, + 0.0037180844228714705, + 0.003670726902782917, + 0.0037492725532501936, + 0.0036270583514124155, + 0.0026501528918743134, + 0.005271510686725378, + 0.004679401405155659, + 0.0047164009883999825, + 0.0052282363176345825, + 0.003070972627028823, + 0.002418231451883912, + 0.0026887645944952965, + 0.0019655260257422924, + 0.004645477049052715, + 0.002523659961298108, + 0.00616505416110158, + 0.008549073711037636, + 0.007208742201328278, + 0.0069155278615653515, + 0.00773165188729763, + 0.00394322769716382, + 0.009060578420758247, + 0.0053628794848918915, + 0.007960312068462372, + 0.019202401861548424, + 0.013985506258904934, + 0.014922826550900936, + 0.02190249413251877, + 0.01218942366540432, + 0.023496689274907112, + 0.014987927861511707, + 0.02114153653383255, + 0.031963545829057693, + 0.04107295721769333, + 0.02972722053527832, + 0.03422994166612625, + 0.014002460986375809, + 0.031489089131355286, + 0.021995510905981064, + 0.03938145563006401, + 0.031091928482055664, + 0.03639329597353935, + 0.032368019223213196, + 0.02901475876569748, + 0.023748736828565598, + 0.046872127801179886, + 0.03741266205906868, + 0.026632603257894516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.3134171962738037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306545481085777 + ], + "opposite_logprob_tokens": [ + -3.1680655479431152 + ], + "logprob_reward": -0.04306545481085777, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306545481085777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12419682741165161, + 0.03755240514874458, + 0.03892958536744118, + 0.03479325398802757, + 0.015781745314598083, + 0.02951890043914318, + 0.06136856973171234, + 0.02824140526354313, + 0.2506891191005707, + 0.37892815470695496 + ], + "credit_map": [ + 0.015173632651567459, + 0.011525201611220837, + 0.018648356199264526, + 0.012603776529431343, + 0.00638601602986455, + 0.023684684187173843, + 0.008525359444320202, + 0.027649812400341034, + 0.005055688321590424, + 0.005044875666499138, + 0.003369406098499894, + 0.006432941183447838, + 0.002754076849669218, + 0.004906254354864359, + 0.004013848956674337, + 0.005975314881652594, + 0.004189709201455116, + 0.0035790421534329653, + 0.004464189056307077, + 0.004268137272447348, + 0.002664142055436969, + 0.009497206658124924, + 0.0030871371272951365, + 0.007180024404078722, + 0.004960371647030115, + 0.0038540188688784838, + 0.005424962844699621, + 0.003892323700711131, + 0.002421317622065544, + 0.004832107573747635, + 0.003623559372499585, + 0.005784595385193825, + 0.0018286422127857804, + 0.002154152374714613, + 0.0016514202579855919, + 0.0018676838371902704, + 0.0011696605943143368, + 0.002449419116601348, + 0.0021569698583334684, + 0.0025037999730557203, + 0.004155443515628576, + 0.0030473333317786455, + 0.002005500253289938, + 0.0027778877411037683, + 0.00250125490128994, + 0.006142543628811836, + 0.0025101483333855867, + 0.006378792691975832, + 0.009218072518706322, + 0.008801808580756187, + 0.009719986468553543, + 0.008309492841362953, + 0.00336736929602921, + 0.009622432291507721, + 0.0046808719635009766, + 0.007648538798093796, + 0.0036921545397490263, + 0.0035632012877613306, + 0.0041749863885343075, + 0.0041520558297634125, + 0.0022652822080999613, + 0.003192683681845665, + 0.002766094636172056, + 0.004434949718415737, + 0.027139738202095032, + 0.04950043931603432, + 0.03732472285628319, + 0.037130191922187805, + 0.010578892193734646, + 0.03169122710824013, + 0.020113077014684677, + 0.03721082583069801, + 0.04341694340109825, + 0.052190735936164856, + 0.0450264997780323, + 0.041179247200489044, + 0.033384982496500015, + 0.07127446681261063, + 0.05294983088970184, + 0.03950547054409981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.04306545481085777 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up depicts a woman on the left side of the screen, her presence imbuing the image with warmth and expressiveness. Her honey-blonde curls fall over the shoulders of a soft, cream-colored cashmere sweater. Warm, natural light from an off-screen source illuminates the right side of her face, highlighting her fair complexion and kind brown eyes. The background to her right is softly out of focus, revealing the blurred shapes of a bookshelf and the gentle glow of a warm, small lamp, creating a cozy and intimate interior. As she speaks, her expression is vivid and captivating; she occasionally blinks slowly, nods thoughtfully, and her gaze briefly shifts to the right of the screen, as if recalling a detail. For emphasis, she deliberately raises her eyebrows, then lets them settle. She then leans forward slightly, a genuine, warm smile spreading across her face as she continues her point, drawing the viewer closer to the conversation. Finally, she straightens her posture, her body relaxing, and her head gently sways from side to side in a comfortable rhythm, conveying a genuine and joyful atmosphere of conversation.", + "seed": 4628, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_11/index_1206_seed_4628_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03387417104095221, + "vlm_logprob": -0.03387417104095221, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.03387417104095221, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06615463644266129 + ], + "opposite_logprob_tokens": [ + -2.816154718399048 + ], + "logprob_reward": -0.06615463644266129, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06615463644266129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1389199048280716, + 0.054006341844797134, + 0.0935976654291153, + 0.05264876037836075, + 0.017173059284687042, + 0.05902258679270744, + 0.03373173251748085, + 0.17614150047302246, + 0.027631225064396858, + 0.3471272587776184 + ], + "credit_map": [ + 0.015376381576061249, + 0.018083667382597923, + 0.01585913635790348, + 0.018342001363635063, + 0.010148277506232262, + 0.016230139881372452, + 0.018847446888685226, + 0.026032838970422745, + 0.005375314969569445, + 0.00610712543129921, + 0.004223513882607222, + 0.004790356382727623, + 0.00605790363624692, + 0.0088104959577322, + 0.00623265840113163, + 0.012408972717821598, + 0.01699194312095642, + 0.014252339489758015, + 0.011964038945734501, + 0.008057591505348682, + 0.010367614217102528, + 0.012876422144472599, + 0.010389154776930809, + 0.00869856495410204, + 0.006390212569385767, + 0.00556462537497282, + 0.010635657235980034, + 0.004782349802553654, + 0.003912937827408314, + 0.006029130890965462, + 0.008568556047976017, + 0.006765285972505808, + 0.0030477282125502825, + 0.0023735978174954653, + 0.0029314791318029165, + 0.0014444147236645222, + 0.0014809928834438324, + 0.0023113673087209463, + 0.0015856116078794003, + 0.0019978682976216078, + 0.007638067938387394, + 0.013341243378818035, + 0.005730141419917345, + 0.007361189927905798, + 0.00563909811899066, + 0.00708432774990797, + 0.007906992919743061, + 0.00432152533903718, + 0.0032063594553619623, + 0.005175638943910599, + 0.0032982989214360714, + 0.0034635993652045727, + 0.008096667937934399, + 0.00483442097902298, + 0.0032135595101863146, + 0.002443190896883607, + 0.020767854526638985, + 0.023554792627692223, + 0.014709973707795143, + 0.015090868808329105, + 0.01864888146519661, + 0.03415583446621895, + 0.02163446508347988, + 0.02757883444428444, + 0.003670554840937257, + 0.0037584747187793255, + 0.003236460965126753, + 0.002719973213970661, + 0.0040289429016411304, + 0.003477606223896146, + 0.0042157466523349285, + 0.0025234674103558064, + 0.014677428640425205, + 0.016415102407336235, + 0.01820283755660057, + 0.017555560916662216, + 0.10716036707162857, + 0.138837069272995, + 0.019279222935438156, + 0.014999653212726116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.06615463644266129 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029863810166716576 + ], + "opposite_logprob_tokens": [ + -3.5298638343811035 + ], + "logprob_reward": -0.029863810166716576, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029863810166716576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16557207703590393, + 0.08441003412008286, + 0.03531000763177872, + 0.06477255374193192, + 0.051177021116018295, + 0.04921111837029457, + 0.09706422686576843, + 0.1105521097779274, + 0.15030598640441895, + 0.19162483513355255 + ], + "credit_map": [ + 0.021672850474715233, + 0.02309035323560238, + 0.02459881082177162, + 0.019191168248653412, + 0.01290012989193201, + 0.02149292081594467, + 0.015590466558933258, + 0.027035387232899666, + 0.009057944640517235, + 0.011281854473054409, + 0.007152548525482416, + 0.009009402245283127, + 0.010355276055634022, + 0.012800371274352074, + 0.007407451514154673, + 0.017345182597637177, + 0.004256805405020714, + 0.005361201241612434, + 0.005007240921258926, + 0.003645421704277396, + 0.0038467992562800646, + 0.003949829842895269, + 0.0044451747089624405, + 0.0047975340858101845, + 0.006638175807893276, + 0.008291005156934261, + 0.014205336570739746, + 0.006424559745937586, + 0.0067538428120315075, + 0.007911907508969307, + 0.007444992195814848, + 0.007102734874933958, + 0.005083946511149406, + 0.006425660103559494, + 0.007930473424494267, + 0.004675732459872961, + 0.0050214785151183605, + 0.006605371832847595, + 0.005639777053147554, + 0.009794577956199646, + 0.008003901690244675, + 0.006033819634467363, + 0.006953977979719639, + 0.00491266418248415, + 0.005855480674654245, + 0.005845047999173403, + 0.005965056829154491, + 0.0056411707773804665, + 0.010754217393696308, + 0.015536841005086899, + 0.012191642075777054, + 0.00991837214678526, + 0.01402505673468113, + 0.01161037478595972, + 0.011597930453717709, + 0.011429796926677227, + 0.011356599628925323, + 0.01593056507408619, + 0.01237714197486639, + 0.012416444718837738, + 0.017053335905075073, + 0.019505824893712997, + 0.013125574216246605, + 0.008786622434854507, + 0.013368501327931881, + 0.026077693328261375, + 0.018010400235652924, + 0.013686374761164188, + 0.02869153581559658, + 0.017927514389157295, + 0.021131915971636772, + 0.011412045918405056, + 0.009832823649048805, + 0.008821764960885048, + 0.011347156949341297, + 0.012002705596387386, + 0.05378716439008713, + 0.07111677527427673, + 0.01415269635617733, + 0.010563747957348824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.029863810166716576 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016084907576441765 + ], + "opposite_logprob_tokens": [ + -4.141084671020508 + ], + "logprob_reward": -0.016084907576441765, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016084907576441765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05297815799713135, + 0.025148171931505203, + 0.03283018246293068, + 0.030560918152332306, + 0.01928349770605564, + 0.11060331016778946, + 0.04094699025154114, + 0.11126989126205444, + 0.10522263497114182, + 0.47115620970726013 + ], + "credit_map": [ + 0.004638139624148607, + 0.006136731244623661, + 0.007346339989453554, + 0.004741549491882324, + 0.004935425706207752, + 0.007353281602263451, + 0.005448541138321161, + 0.012378147803246975, + 0.002831302350386977, + 0.003551982343196869, + 0.0022321981377899647, + 0.0025018576998263597, + 0.0035818142350763083, + 0.0035076658241450787, + 0.002550708595663309, + 0.004390643443912268, + 0.004291799385100603, + 0.004448493476957083, + 0.0032530655153095722, + 0.0029885005205869675, + 0.002622402273118496, + 0.005277819000184536, + 0.0038616848178207874, + 0.006086415145546198, + 0.004250435158610344, + 0.0027371717151254416, + 0.0074194734916090965, + 0.0032609249465167522, + 0.002534235594794154, + 0.0034928901586681604, + 0.004151234868913889, + 0.002714550821110606, + 0.002814213279634714, + 0.0021072064992040396, + 0.0021899226121604443, + 0.001619416056200862, + 0.0013907921966165304, + 0.002884146524593234, + 0.0015461994335055351, + 0.0047315992414951324, + 0.016232872381806374, + 0.028183134272694588, + 0.011293040588498116, + 0.010384732857346535, + 0.01037604920566082, + 0.015995223075151443, + 0.008490409702062607, + 0.009647846221923828, + 0.0052803256548941135, + 0.009223449975252151, + 0.004739255178719759, + 0.004118140786886215, + 0.006054065655916929, + 0.005230884067714214, + 0.003487854963168502, + 0.0028130130376666784, + 0.017762495204806328, + 0.013674970716238022, + 0.006822606083005667, + 0.008742006495594978, + 0.010273978114128113, + 0.0363633893430233, + 0.0111808767542243, + 0.006449565757066011, + 0.009633781388401985, + 0.014710522256791592, + 0.008485604077577591, + 0.01059731189161539, + 0.027244163677096367, + 0.011335198767483234, + 0.015210269019007683, + 0.008005787618458271, + 0.020615922287106514, + 0.020737357437610626, + 0.021881919354200363, + 0.023827925324440002, + 0.13982364535331726, + 0.18249712884426117, + 0.03793822601437569, + 0.023834096267819405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.016084907576441765 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306385666131973 + ], + "opposite_logprob_tokens": [ + -3.1680638790130615 + ], + "logprob_reward": -0.04306385666131973, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306385666131973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08068517595529556, + 0.05711968615651131, + 0.031457673758268356, + 0.03187284991145134, + 0.021709425374865532, + 0.18905724585056305, + 0.12224584817886353, + 0.10854309797286987, + 0.13581298291683197, + 0.22149606049060822 + ], + "credit_map": [ + 0.012171948328614235, + 0.009106623940169811, + 0.01078637782484293, + 0.008585221134126186, + 0.007005379535257816, + 0.009867683984339237, + 0.010754348710179329, + 0.012407590635120869, + 0.0071817259304225445, + 0.005252247676253319, + 0.006663977634161711, + 0.0048295664601027966, + 0.008586391806602478, + 0.008770979940891266, + 0.008187726140022278, + 0.007647067774087191, + 0.004795799031853676, + 0.004127355758100748, + 0.00308835762552917, + 0.00277568562887609, + 0.003990002442151308, + 0.0036839928943663836, + 0.0035398234613239765, + 0.0054566566832363605, + 0.003901716787368059, + 0.004458988551050425, + 0.0036461716517806053, + 0.004208601079881191, + 0.0028263588901609182, + 0.0048187850043177605, + 0.004153104033321142, + 0.0038591257762163877, + 0.0032956514041870832, + 0.0029730121605098248, + 0.0027466854080557823, + 0.0024188512470573187, + 0.0021400407422333956, + 0.002919211285188794, + 0.0020769701804965734, + 0.003139004809781909, + 0.021656809374690056, + 0.05388442054390907, + 0.01577826589345932, + 0.018984610214829445, + 0.021735621616244316, + 0.028148271143436432, + 0.016852056607604027, + 0.012017176486551762, + 0.009679550305008888, + 0.020853372290730476, + 0.00926031544804573, + 0.010943278670310974, + 0.03102746233344078, + 0.017177117988467216, + 0.016395725309848785, + 0.006909020710736513, + 0.015535343438386917, + 0.012595204636454582, + 0.006875073071569204, + 0.010996910743415356, + 0.009996501728892326, + 0.03142838552594185, + 0.012876385822892189, + 0.008239293470978737, + 0.012271606363356113, + 0.03291754797101021, + 0.012728344649076462, + 0.019473489373922348, + 0.019474275410175323, + 0.018291030079126358, + 0.010969741269946098, + 0.009686951525509357, + 0.012198295444250107, + 0.011169645935297012, + 0.01309740636497736, + 0.011564575135707855, + 0.06795529276132584, + 0.08578956872224808, + 0.010923963040113449, + 0.008797314018011093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.04306385666131973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420364435762167 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.01420364435762167, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420364435762167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050356149673461914, + 0.02202332764863968, + 0.023194335401058197, + 0.018578873947262764, + 0.020082755014300346, + 0.14212888479232788, + 0.038004010915756226, + 0.0861160159111023, + 0.046594131737947464, + 0.55292147397995 + ], + "credit_map": [ + 0.004447723738849163, + 0.00789465382695198, + 0.006664396729320288, + 0.006366012617945671, + 0.00609064195305109, + 0.007072790525853634, + 0.005731041543185711, + 0.006088887341320515, + 0.0027893208898603916, + 0.0034580936189740896, + 0.0027208058163523674, + 0.0026582335121929646, + 0.002132764784619212, + 0.002900942461565137, + 0.0021898550912737846, + 0.00317331007681787, + 0.002704707905650139, + 0.003185332752764225, + 0.0026737486477941275, + 0.0018864935263991356, + 0.002151529537513852, + 0.0030784106347709894, + 0.0030739353969693184, + 0.0044401781633496284, + 0.0023310023825615644, + 0.0026378585025668144, + 0.001747881411574781, + 0.002231735037639737, + 0.001983222784474492, + 0.0021361622493714094, + 0.0021566760260611773, + 0.0033543347381055355, + 0.003005230100825429, + 0.002132000168785453, + 0.0019513566512614489, + 0.0017582174623385072, + 0.0014087557792663574, + 0.0031856175046414137, + 0.0012364600552245975, + 0.005405118688941002, + 0.015080979093909264, + 0.03564959019422531, + 0.011614777147769928, + 0.01628861203789711, + 0.016563817858695984, + 0.02148962765932083, + 0.014400218613445759, + 0.011041264981031418, + 0.004059792496263981, + 0.007750384509563446, + 0.003982532769441605, + 0.003914565779268742, + 0.006714034825563431, + 0.004984972067177296, + 0.003797263139858842, + 0.0028004643972963095, + 0.01146948616951704, + 0.011415551416575909, + 0.0070898327976465225, + 0.007732301019132137, + 0.007694656029343605, + 0.02394036203622818, + 0.009495560079813004, + 0.007278267294168472, + 0.003433572594076395, + 0.010705817490816116, + 0.004970226436853409, + 0.005603952798992395, + 0.0072371577844023705, + 0.006209738086909056, + 0.005884276237338781, + 0.002549390774220228, + 0.03065587766468525, + 0.02848861552774906, + 0.040392547845840454, + 0.035904865711927414, + 0.16898371279239655, + 0.20786894857883453, + 0.023868262767791748, + 0.016758689656853676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.01420364435762167 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static close-up depicts a woman on the left side of the screen, her presence imbuing the image with warmth and expressiveness. Her honey-blonde curls fall over the shoulders of a soft, cream-colored cashmere sweater. Warm, natural light from an off-screen source illuminates the right side of her face, highlighting her fair complexion and kind brown eyes. The background to her right is softly out of focus, revealing the blurred shapes of a bookshelf and the gentle glow of a warm, small lamp, creating a cozy and intimate interior. As she speaks, her expression is vivid and captivating; she occasionally blinks slowly, nods thoughtfully, and her gaze briefly shifts to the right of the screen, as if recalling a detail. For emphasis, she deliberately raises her eyebrows, then lets them settle. She then leans forward slightly, a genuine, warm smile spreading across her face as she continues her point, drawing the viewer closer to the conversation. Finally, she straightens her posture, her body relaxing, and her head gently sways from side to side in a comfortable rhythm, conveying a genuine and joyful atmosphere of conversation.", + "seed": 4629, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_11/index_1206_seed_4629_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05433712378144264, + "vlm_logprob": -0.05433712378144264, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05433712378144264, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011512276716530323 + ], + "opposite_logprob_tokens": [ + -4.636512279510498 + ], + "logprob_reward": -0.011512276716530323, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011512276716530323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2721084654331207, + 0.05187520757317543, + 0.09253998100757599, + 0.03241106495261192, + 0.12310238927602768, + 0.03593656048178673, + 0.04079994931817055, + 0.046911899000406265, + 0.1566152423620224, + 0.14769917726516724 + ], + "credit_map": [ + 0.0281162541359663, + 0.029000254347920418, + 0.027815070003271103, + 0.03615808114409447, + 0.04154575243592262, + 0.03003389947116375, + 0.05453307181596756, + 0.024906082078814507, + 0.004696611315011978, + 0.008393475785851479, + 0.004995808936655521, + 0.006700415629893541, + 0.010283077135682106, + 0.005792734678834677, + 0.0056318254210054874, + 0.005381257273256779, + 0.008048814721405506, + 0.010868681594729424, + 0.0061086565256118774, + 0.01667105220258236, + 0.013104556128382683, + 0.009602873586118221, + 0.013483780436217785, + 0.014651564881205559, + 0.00347006693482399, + 0.0041834511794149876, + 0.003057045629248023, + 0.005687653552740812, + 0.005578177981078625, + 0.004304773639887571, + 0.004095797426998615, + 0.0020340988412499428, + 0.013996616937220097, + 0.011708339676260948, + 0.010216666385531425, + 0.01770283654332161, + 0.017327893525362015, + 0.013231311924755573, + 0.025412127375602722, + 0.013506588526070118, + 0.003941560164093971, + 0.004026533104479313, + 0.0021176219452172518, + 0.005935709457844496, + 0.006101981271058321, + 0.0038837867323309183, + 0.005970520433038473, + 0.003958850167691708, + 0.0046892655082046986, + 0.004433566238731146, + 0.006730434484779835, + 0.0052865296602249146, + 0.007604396436363459, + 0.004022217355668545, + 0.006209385581314564, + 0.0018241516081616282, + 0.0043259793892502785, + 0.006008444353938103, + 0.004234653431922197, + 0.006534615531563759, + 0.008551366627216339, + 0.0058281621895730495, + 0.007625878788530827, + 0.0038027972914278507, + 0.015060563571751118, + 0.028101827949285507, + 0.012880222871899605, + 0.019277183338999748, + 0.021544933319091797, + 0.018500404432415962, + 0.018080422654747963, + 0.023169683292508125, + 0.020111970603466034, + 0.016885604709386826, + 0.011695098131895065, + 0.017207153141498566, + 0.01930423453450203, + 0.030788971111178398, + 0.010322260670363903, + 0.021383892744779587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.011512276716530323 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03376022353768349 + ], + "opposite_logprob_tokens": [ + -3.4087603092193604 + ], + "logprob_reward": -0.03376022353768349, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03376022353768349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26367419958114624, + 0.11892225593328476, + 0.0788138210773468, + 0.045038409531116486, + 0.11511828750371933, + 0.05277438834309578, + 0.043580591678619385, + 0.04091297462582588, + 0.09345733374357224, + 0.1477077305316925 + ], + "credit_map": [ + 0.0357719324529171, + 0.030328551307320595, + 0.02957374043762684, + 0.03241214156150818, + 0.04804356023669243, + 0.02432488091289997, + 0.041934795677661896, + 0.02128458581864834, + 0.014576052315533161, + 0.015140092931687832, + 0.011562030762434006, + 0.017587922513484955, + 0.01668882742524147, + 0.02010822482407093, + 0.015123910270631313, + 0.00813520047813654, + 0.011073313653469086, + 0.007771117612719536, + 0.0086024459451437, + 0.011679164133965969, + 0.014390972442924976, + 0.009875227697193623, + 0.010452326387166977, + 0.004969256464391947, + 0.004760988522320986, + 0.005097229965031147, + 0.004935362841933966, + 0.005995703861117363, + 0.00800609216094017, + 0.004915305413305759, + 0.00771383848041296, + 0.0036138861905783415, + 0.015594640746712685, + 0.013302509672939777, + 0.021945951506495476, + 0.014778356067836285, + 0.015867037698626518, + 0.014953835867345333, + 0.011608910746872425, + 0.0070670428685843945, + 0.009482371620833874, + 0.0060356720350682735, + 0.005605512298643589, + 0.008547324687242508, + 0.00663131196051836, + 0.005834386218339205, + 0.006868473254144192, + 0.003769338596612215, + 0.0033249163534492254, + 0.004861920140683651, + 0.004937830381095409, + 0.0054930513724684715, + 0.008035043254494667, + 0.0049764234572649, + 0.009102392941713333, + 0.002849013777449727, + 0.0040641063824296, + 0.0052595557644963264, + 0.0035171760246157646, + 0.006945843808352947, + 0.006310228258371353, + 0.005036048591136932, + 0.007275683805346489, + 0.0025043312925845385, + 0.007871151901781559, + 0.017434317618608475, + 0.009487839415669441, + 0.012562266550958157, + 0.013087045401334763, + 0.010614458471536636, + 0.009805134497582912, + 0.012595119886100292, + 0.01974758319556713, + 0.020472384989261627, + 0.019966036081314087, + 0.016106609255075455, + 0.021172629669308662, + 0.01802908629179001, + 0.0138460174202919, + 0.01836736500263214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.03376022353768349 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011087591759860516 + ], + "opposite_logprob_tokens": [ + -4.511087417602539 + ], + "logprob_reward": -0.011087591759860516, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011087591759860516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16412867605686188, + 0.09462715685367584, + 0.06666527688503265, + 0.05450815707445145, + 0.04816501960158348, + 0.07382013648748398, + 0.01726226694881916, + 0.04269620403647423, + 0.11943159252405167, + 0.3186955153942108 + ], + "credit_map": [ + 0.01638740301132202, + 0.014378017745912075, + 0.01256672665476799, + 0.019938534125685692, + 0.02243083342909813, + 0.015599836595356464, + 0.02169233188033104, + 0.04113499075174332, + 0.015699608251452446, + 0.013457429595291615, + 0.006951690185815096, + 0.011334924958646297, + 0.011916576884686947, + 0.011674982495605946, + 0.009794211015105247, + 0.013797733932733536, + 0.0048415628261864185, + 0.009417178109288216, + 0.0041687386110424995, + 0.012889106757938862, + 0.010679533705115318, + 0.00724510382860899, + 0.00878910068422556, + 0.008634955622255802, + 0.005043003708124161, + 0.005693456623703241, + 0.003853890113532543, + 0.008006119169294834, + 0.007735785096883774, + 0.005481748841702938, + 0.00920880027115345, + 0.009485348127782345, + 0.0054088919423520565, + 0.005125031806528568, + 0.004812346771359444, + 0.00549291679635644, + 0.007974707521498203, + 0.0053461831994354725, + 0.008883357979357243, + 0.005121584050357342, + 0.008625848218798637, + 0.008466369472444057, + 0.005502050276845694, + 0.010677934624254704, + 0.014024038799107075, + 0.007613350171595812, + 0.013152853585779667, + 0.005757689010351896, + 0.0018666833639144897, + 0.002111677313223481, + 0.0017986255697906017, + 0.002018080558627844, + 0.0030494059901684523, + 0.0018375685904175043, + 0.0033956205006688833, + 0.0011846049455925822, + 0.006483879406005144, + 0.0051563275046646595, + 0.0036752596497535706, + 0.005063993390649557, + 0.006756357848644257, + 0.005388123914599419, + 0.006277348846197128, + 0.0038949125446379185, + 0.009928766638040543, + 0.018629256635904312, + 0.010957622900605202, + 0.01453737635165453, + 0.01514514721930027, + 0.013286587782204151, + 0.013399029150605202, + 0.023547803983092308, + 0.02971169725060463, + 0.02763323299586773, + 0.020186495035886765, + 0.023263446986675262, + 0.046355292201042175, + 0.06938953697681427, + 0.024655116721987724, + 0.07750070095062256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.011087591759860516 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603633314371109 + ], + "opposite_logprob_tokens": [ + -1.9103633165359497 + ], + "logprob_reward": -0.1603633314371109, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603633314371109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09470316022634506, + 0.34199443459510803, + 0.06575939804315567, + 0.026670778170228004, + 0.04748976603150368, + 0.04055854678153992, + 0.03228845074772835, + 0.032691460102796555, + 0.09250844269990921, + 0.22533553838729858 + ], + "credit_map": [ + 0.007660832721740007, + 0.009350597858428955, + 0.006801220588386059, + 0.013542495667934418, + 0.019067751243710518, + 0.009095052257180214, + 0.02118571847677231, + 0.00799949187785387, + 0.04932308942079544, + 0.039957765489816666, + 0.02181978151202202, + 0.04308946803212166, + 0.056939318776130676, + 0.04044124111533165, + 0.05304834619164467, + 0.03737540543079376, + 0.010034027509391308, + 0.006024286150932312, + 0.005329555831849575, + 0.009217975661158562, + 0.01063681859523058, + 0.0066717155277729034, + 0.010682792402803898, + 0.00716222682967782, + 0.0022326072212308645, + 0.0036425821017473936, + 0.0019293463556095958, + 0.0038191431667655706, + 0.004963813349604607, + 0.0028814696706831455, + 0.004889894276857376, + 0.0023119207471609116, + 0.004830782767385244, + 0.003931851126253605, + 0.0031362329609692097, + 0.006877796724438667, + 0.006357809528708458, + 0.004505332559347153, + 0.008920084685087204, + 0.008929876610636711, + 0.005207081325352192, + 0.004851481877267361, + 0.0037290717009454966, + 0.0055261217057704926, + 0.0061292205937206745, + 0.004197901580482721, + 0.006477543618530035, + 0.0044401297345757484, + 0.0030016563832759857, + 0.00272618280723691, + 0.0019212698098272085, + 0.004625412635505199, + 0.005156819708645344, + 0.003783280961215496, + 0.007616122718900442, + 0.0034577040933072567, + 0.0028682691045105457, + 0.003848322667181492, + 0.002776120323687792, + 0.005363043397665024, + 0.0052729882299900055, + 0.0034484502393752337, + 0.0072048441506922245, + 0.0019094225717708468, + 0.008438380435109138, + 0.01757802441716194, + 0.007423098664730787, + 0.011367528699338436, + 0.012555314227938652, + 0.010162301361560822, + 0.008866449818015099, + 0.01611735112965107, + 0.02891450934112072, + 0.04469514265656471, + 0.02827010490000248, + 0.02832282893359661, + 0.02796720527112484, + 0.02543785609304905, + 0.027253370732069016, + 0.014474539086222649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.1603633314371109 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a woman positioned on the left side of the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a soft, cream-colored cashmere sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background to the right of the woman softly out of focus, showing a bookshelf and a small lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman raise her eyebrows for emphasis and then let them settle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman lean forward slightly and smile genuinely before straightening her posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054962195456027985 + ], + "opposite_logprob_tokens": [ + -2.929962158203125 + ], + "logprob_reward": -0.054962195456027985, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054962195456027985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10597994178533554, + 0.26631787419319153, + 0.12345811724662781, + 0.021503502503037453, + 0.02059995010495186, + 0.028474420309066772, + 0.06280525028705597, + 0.053325630724430084, + 0.06553244590759277, + 0.2520028352737427 + ], + "credit_map": [ + 0.01035398617386818, + 0.010145171545445919, + 0.006053386256098747, + 0.015599382109940052, + 0.02107764407992363, + 0.010191037319600582, + 0.026614287868142128, + 0.005945053417235613, + 0.039017193019390106, + 0.02410801686346531, + 0.016601750627160072, + 0.03466641530394554, + 0.039890825748443604, + 0.02564050443470478, + 0.06365145742893219, + 0.022741731256246567, + 0.019252674654126167, + 0.011321185156702995, + 0.008151419460773468, + 0.01847536489367485, + 0.0184478797018528, + 0.0115116648375988, + 0.020913327112793922, + 0.015384609811007977, + 0.002577001228928566, + 0.0022817945573478937, + 0.0017669270746409893, + 0.003623769385740161, + 0.0036370318848639727, + 0.0020496624056249857, + 0.004449247848242521, + 0.0011180713772773743, + 0.002109834458678961, + 0.0021731264423578978, + 0.0017295238794758916, + 0.0032466582488268614, + 0.002376243472099304, + 0.0018470885697752237, + 0.005357659887522459, + 0.0017598173581063747, + 0.0028749401681125164, + 0.004192691762000322, + 0.0019472027197480202, + 0.0040384866297245026, + 0.004224001429975033, + 0.0026344696525484324, + 0.004567132797092199, + 0.003995496779680252, + 0.005989708472043276, + 0.006293103098869324, + 0.0035519327502697706, + 0.007639244664460421, + 0.011830689385533333, + 0.007074754219502211, + 0.015318403020501137, + 0.005107414908707142, + 0.0048010158352553844, + 0.007738702930510044, + 0.004072967451065779, + 0.008205828256905079, + 0.0072999210096895695, + 0.005475591402500868, + 0.011217313818633556, + 0.004514296073466539, + 0.004748817998915911, + 0.012887225486338139, + 0.0055746412836015224, + 0.008852604776620865, + 0.010063006542623043, + 0.006811321713030338, + 0.006789869628846645, + 0.009804962202906609, + 0.025466926395893097, + 0.03745373338460922, + 0.04392128437757492, + 0.030032338574528694, + 0.03292335197329521, + 0.02478496916592121, + 0.040425948798656464, + 0.016994275152683258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.054962195456027985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_12/index_tensor([707, 707])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_12/index_tensor([707, 707])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..139ee599fe43e502ce5fd997f8bf1f8f3d9b7a43 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_12/index_tensor([707, 707])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a slightly high-angle medium close-up, the camera smoothly pans to the right, its movements perfectly synchronized with the meticulous movements of a seamstress working in her sunlit studio. The central subject is her hands, one holding a piece of emerald green silk flat on the surface of a worn, light oak cutting table, the other guiding a pair of heavy, gleaming chrome-plated tailor's scissors. Seen from the side, she is an elderly woman, her silver hair tied in a neat bun, her face a look of intense concentration. The scissors move steadily and deliberately, the polished steel blades opening and closing with a satisfyingly crisp rhythm along a faint chalk line, parting the uneven fabric with unwavering precision. The scene is illuminated by soft natural light streaming in from the large windows, creating a gentle sheen on the silk and highlighting the fine grain of the wooden table. In the softly blurred background, colorful fabric rolls, spools, and an old-fashioned sewing machine create a rich, textured backdrop. The atmosphere is one of quiet concentration and exquisite craftsmanship, a serene and intimate portrait of an artisan lost in her craft.", + "seed": 506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_12/index_707_seed_506_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0836401553824544, + "vlm_logprob": -0.0836401553824544, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.0836401553824544, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08650598675012589 + ], + "opposite_logprob_tokens": [ + -2.586505889892578 + ], + "logprob_reward": -0.08650598675012589, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08650598675012589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09133771061897278, + 0.1638328582048416, + 0.047850366681814194, + 0.06006334722042084, + 0.08398061245679855, + 0.1603725254535675, + 0.3088501989841461, + 0.019648047164082527, + 0.03492658585309982, + 0.02913770265877247 + ], + "credit_map": [ + 0.011886884458363056, + 0.011054004542529583, + 0.007965711876749992, + 0.011973354034125805, + 0.015953708440065384, + 0.009463547728955746, + 0.014958463609218597, + 0.008082046173512936, + 0.014234607107937336, + 0.0073763905093073845, + 0.055922508239746094, + 0.006447117310017347, + 0.007755196653306484, + 0.010679611004889011, + 0.04408090189099312, + 0.017336530610919, + 0.005465747322887182, + 0.004116316791623831, + 0.004938996862620115, + 0.0031270685140043497, + 0.004382563289254904, + 0.008750369772315025, + 0.006296258885413408, + 0.010773051530122757, + 0.006440842989832163, + 0.016887517645955086, + 0.006659294478595257, + 0.004626988433301449, + 0.005251593422144651, + 0.00698717450723052, + 0.0050540463998913765, + 0.00815589539706707, + 0.009642647579312325, + 0.015597714111208916, + 0.011273497715592384, + 0.005260412115603685, + 0.006494245491921902, + 0.015214300714433193, + 0.007991187274456024, + 0.01250661350786686, + 0.01536505576223135, + 0.01472207996994257, + 0.009173000231385231, + 0.007549350615590811, + 0.014114618301391602, + 0.031341247260570526, + 0.041738998144865036, + 0.02636818401515484, + 0.01649138145148754, + 0.016238799318671227, + 0.009235749952495098, + 0.009421913884580135, + 0.01930941827595234, + 0.024451056495308876, + 0.10883692651987076, + 0.10486500710248947, + 0.0033452094066888094, + 0.00235787034034729, + 0.0025288898032158613, + 0.0017081922851502895, + 0.002478820038959384, + 0.002960050478577614, + 0.0013491511344909668, + 0.002919865073636174, + 0.006349556613713503, + 0.0033274609595537186, + 0.002807382494211197, + 0.0020269015803933144, + 0.0027709659188985825, + 0.006532407831400633, + 0.004296314436942339, + 0.00681559881195426, + 0.0034188362769782543, + 0.002423250349238515, + 0.0018269405700266361, + 0.0015101501485332847, + 0.006608871277421713, + 0.006356126628816128, + 0.003395285690203309, + 0.003598242998123169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.08650598675012589 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006758331321179867 + ], + "opposite_logprob_tokens": [ + -5.006758213043213 + ], + "logprob_reward": -0.006758331321179867, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006758331321179867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0584256537258625, + 0.39516890048980713, + 0.09591606259346008, + 0.07305977493524551, + 0.10104919224977493, + 0.08904767781496048, + 0.09438889473676682, + 0.04283153638243675, + 0.024926770478487015, + 0.025185495615005493 + ], + "credit_map": [ + 0.007217158563435078, + 0.0061648222617805, + 0.005597339943051338, + 0.009672557935118675, + 0.009401780553162098, + 0.005278440658003092, + 0.009726384654641151, + 0.005367174278944731, + 0.03199471905827522, + 0.014430982060730457, + 0.13534016907215118, + 0.013507083989679813, + 0.017716296017169952, + 0.02351079136133194, + 0.1313953846693039, + 0.02727351151406765, + 0.008455788716673851, + 0.008960176259279251, + 0.03449004888534546, + 0.006205280311405659, + 0.008511045016348362, + 0.010489516891539097, + 0.006955291144549847, + 0.011848913505673409, + 0.007288812659680843, + 0.012825891375541687, + 0.01068782713264227, + 0.008572075515985489, + 0.006555598694831133, + 0.007532758638262749, + 0.006847355514764786, + 0.012749460525810719, + 0.009262952953577042, + 0.020594103261828423, + 0.01866193488240242, + 0.012711573392152786, + 0.008248288184404373, + 0.01146238949149847, + 0.009295646101236343, + 0.010812311433255672, + 0.0074341874569654465, + 0.007503045257180929, + 0.008721956983208656, + 0.007653386797755957, + 0.0052268183790147305, + 0.008819809183478355, + 0.02537352219223976, + 0.01831495761871338, + 0.00442851847037673, + 0.0037170243449509144, + 0.004614606965333223, + 0.004589793737977743, + 0.005984626244753599, + 0.006606854032725096, + 0.034695230424404144, + 0.029752256348729134, + 0.005194083321839571, + 0.00526995025575161, + 0.0039714304730296135, + 0.003084850497543812, + 0.004916137550026178, + 0.01056490745395422, + 0.004351345356553793, + 0.00547883752733469, + 0.005304086487740278, + 0.0025709220208227634, + 0.003067213110625744, + 0.0022124038077890873, + 0.002342396415770054, + 0.002646960085257888, + 0.0033321548253297806, + 0.0034506346564739943, + 0.005790635012090206, + 0.0033675432205200195, + 0.002218412235379219, + 0.0013951126020401716, + 0.0025185225531458855, + 0.004149257205426693, + 0.002395824994891882, + 0.0033501889556646347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.006758331321179867 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430685393512249 + ], + "opposite_logprob_tokens": [ + -3.1680686473846436 + ], + "logprob_reward": -0.0430685393512249, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430685393512249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.146833598613739, + 0.23916815221309662, + 0.16429533064365387, + 0.058458298444747925, + 0.10260048508644104, + 0.06102067977190018, + 0.04496253654360771, + 0.053727731108665466, + 0.09816864877939224, + 0.030764516443014145 + ], + "credit_map": [ + 0.011057284660637379, + 0.01530052162706852, + 0.010173982940614223, + 0.012779941782355309, + 0.02671879157423973, + 0.014327939599752426, + 0.04033881425857544, + 0.016136327758431435, + 0.02011883817613125, + 0.009769359603524208, + 0.07099524140357971, + 0.008774757385253906, + 0.010792912915349007, + 0.016392797231674194, + 0.0844750925898552, + 0.01784915290772915, + 0.013607637025415897, + 0.013805885799229145, + 0.05686106160283089, + 0.010969842784106731, + 0.012082982808351517, + 0.02351144142448902, + 0.011947860941290855, + 0.021508613601326942, + 0.0061174132861196995, + 0.013583202846348286, + 0.007202562410384417, + 0.005138171371072531, + 0.0057408506982028484, + 0.008373119868338108, + 0.005778053775429726, + 0.0065249232575297356, + 0.013028434477746487, + 0.009577552787959576, + 0.02236831746995449, + 0.012592175975441933, + 0.00728801591321826, + 0.01350878830999136, + 0.009262265637516975, + 0.014974942430853844, + 0.006304826587438583, + 0.007334183435887098, + 0.0057712518610060215, + 0.0038044475950300694, + 0.006308961659669876, + 0.013162042014300823, + 0.010190670378506184, + 0.008144299499690533, + 0.0045488569885492325, + 0.004483092110604048, + 0.004536747001111507, + 0.0033084829337894917, + 0.006755598820745945, + 0.007841614075005054, + 0.008317863568663597, + 0.005170279648154974, + 0.009031666442751884, + 0.006343713961541653, + 0.007177798077464104, + 0.005617858376353979, + 0.004280895460397005, + 0.008193357847630978, + 0.005656090565025806, + 0.0074263522401452065, + 0.013122721575200558, + 0.009590232744812965, + 0.008572729304432869, + 0.007091497536748648, + 0.01419647503644228, + 0.012748399749398232, + 0.01835835725069046, + 0.01448824256658554, + 0.003368409350514412, + 0.0024824475403875113, + 0.003580409334972501, + 0.0020703317131847143, + 0.004292410798370838, + 0.007014768198132515, + 0.0035265099722892046, + 0.004429229069501162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0430685393512249 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980896644294262 + ], + "opposite_logprob_tokens": [ + -3.52980899810791 + ], + "logprob_reward": -0.02980896644294262, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980896644294262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07222016900777817, + 0.43488457798957825, + 0.08701299875974655, + 0.05190487205982208, + 0.04079676792025566, + 0.0725276842713356, + 0.1389334797859192, + 0.033288538455963135, + 0.052818071097135544, + 0.015612825751304626 + ], + "credit_map": [ + 0.01371674332767725, + 0.009512094780802727, + 0.007498977240175009, + 0.006080720108002424, + 0.010077493265271187, + 0.00796353816986084, + 0.008426476269960403, + 0.008944131433963776, + 0.035462379455566406, + 0.014273158274590969, + 0.1374778300523758, + 0.013950477354228497, + 0.017557142302393913, + 0.027569210156798363, + 0.15667937695980072, + 0.031915049999952316, + 0.008277577348053455, + 0.008344494737684727, + 0.02760487049818039, + 0.005385949742048979, + 0.006060723215341568, + 0.010875011794269085, + 0.012514723464846611, + 0.007949655875563622, + 0.006061846390366554, + 0.01348025631159544, + 0.006033611483871937, + 0.005602302495390177, + 0.004591007251292467, + 0.0057482048869132996, + 0.004238275345414877, + 0.0061493683606386185, + 0.004995489027351141, + 0.003443346591666341, + 0.004190464969724417, + 0.003996902145445347, + 0.003188807051628828, + 0.00589712243527174, + 0.006866873241961002, + 0.00821776408702135, + 0.00724497064948082, + 0.006428456865251064, + 0.007773245219141245, + 0.00599614018574357, + 0.005681816954165697, + 0.00984408799558878, + 0.01779627613723278, + 0.011762703768908978, + 0.007319917902350426, + 0.004915810655802488, + 0.003996196668595076, + 0.004658574238419533, + 0.006642872933298349, + 0.0074020689353346825, + 0.05427674576640129, + 0.04972130432724953, + 0.005859061144292355, + 0.004108961205929518, + 0.004231120925396681, + 0.003725931514054537, + 0.003151270095258951, + 0.004001161549240351, + 0.0035739794839173555, + 0.004637053702026606, + 0.0056251781061291695, + 0.004573459271341562, + 0.006949557922780514, + 0.004825557116419077, + 0.0059844814240932465, + 0.007579758297652006, + 0.010751969181001186, + 0.006528113037347794, + 0.001622612588107586, + 0.0010049764532595873, + 0.0018733053002506495, + 0.0012241031508892775, + 0.0021149783860892057, + 0.003821920370683074, + 0.0019827133510261774, + 0.0019682180136442184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.02980896644294262 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520589530467987 + ], + "opposite_logprob_tokens": [ + -1.502058982849121 + ], + "logprob_reward": -0.2520589530467987, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520589530467987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06763961166143417, + 0.6271241903305054, + 0.022609328851103783, + 0.02442316897213459, + 0.037652451545000076, + 0.04028646647930145, + 0.10100611299276352, + 0.012309740297496319, + 0.04553616791963577, + 0.02141270413994789 + ], + "credit_map": [ + 0.014466558583080769, + 0.01010169181972742, + 0.005661400035023689, + 0.006587168201804161, + 0.007184100802987814, + 0.007441755849868059, + 0.0054451702162623405, + 0.010751773603260517, + 0.05162811651825905, + 0.01984158530831337, + 0.17952804267406464, + 0.019395342096686363, + 0.02249889075756073, + 0.04031731188297272, + 0.2383015900850296, + 0.05561333894729614, + 0.002432197332382202, + 0.0017157030524685979, + 0.0035701789893209934, + 0.0030618696473538876, + 0.0017493257764726877, + 0.004423191305249929, + 0.0035659661516547203, + 0.002090900670737028, + 0.0030062526930123568, + 0.005726272705942392, + 0.0034405849874019623, + 0.0023246887139976025, + 0.0023242710158228874, + 0.0029157509561628103, + 0.0019764432217925787, + 0.0027089084032922983, + 0.003857990959659219, + 0.009947888553142548, + 0.0032249605283141136, + 0.0023425479885190725, + 0.003293633460998535, + 0.004925415851175785, + 0.0034342240542173386, + 0.006625796668231487, + 0.0035558694507926702, + 0.0038209122139960527, + 0.004567692521959543, + 0.0037257198709994555, + 0.002241355599835515, + 0.004229072481393814, + 0.010681286454200745, + 0.007464561611413956, + 0.006264207419008017, + 0.0036866480950266123, + 0.004722984973341227, + 0.003598783165216446, + 0.0051199025474488735, + 0.005689556710422039, + 0.036527492105960846, + 0.03539654612541199, + 0.0023938214872032404, + 0.0014426081907004118, + 0.001047909609042108, + 0.0012650479329749942, + 0.0013404779601842165, + 0.0014086441369727254, + 0.0014746092492714524, + 0.0019366226624697447, + 0.006008449476212263, + 0.004782314877957106, + 0.006088880356401205, + 0.0037164061795920134, + 0.004513378720730543, + 0.004942145198583603, + 0.008115570992231369, + 0.007369024679064751, + 0.002910797018557787, + 0.0015802666312083602, + 0.0014026196440681815, + 0.001022468088194728, + 0.004619212821125984, + 0.004635551944375038, + 0.002322195563465357, + 0.0029195949900895357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.2520589530467987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a slightly high-angle medium close-up, the camera smoothly pans to the right, its movements perfectly synchronized with the meticulous movements of a seamstress working in her sunlit studio. The central subject is her hands, one holding a piece of emerald green silk flat on the surface of a worn, light oak cutting table, the other guiding a pair of heavy, gleaming chrome-plated tailor's scissors. Seen from the side, she is an elderly woman, her silver hair tied in a neat bun, her face a look of intense concentration. The scissors move steadily and deliberately, the polished steel blades opening and closing with a satisfyingly crisp rhythm along a faint chalk line, parting the uneven fabric with unwavering precision. The scene is illuminated by soft natural light streaming in from the large windows, creating a gentle sheen on the silk and highlighting the fine grain of the wooden table. In the softly blurred background, colorful fabric rolls, spools, and an old-fashioned sewing machine create a rich, textured backdrop. The atmosphere is one of quiet concentration and exquisite craftsmanship, a serene and intimate portrait of an artisan lost in her craft.", + "seed": 507, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_12/index_707_seed_507_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09545057527720928, + "vlm_logprob": -0.09545057527720928, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.09545057527720928, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16478261351585388 + ], + "opposite_logprob_tokens": [ + -1.9147826433181763 + ], + "logprob_reward": -0.16478261351585388, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16478261351585388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26262304186820984, + 0.19512419402599335, + 0.042434219270944595, + 0.03274009749293327, + 0.0805535539984703, + 0.12462394684553146, + 0.031218040734529495, + 0.06150982156395912, + 0.047960806638002396, + 0.12121222168207169 + ], + "credit_map": [ + 0.03493427485227585, + 0.023920929059386253, + 0.03284512087702751, + 0.052656855434179306, + 0.021031148731708527, + 0.039737142622470856, + 0.024377459660172462, + 0.033120136708021164, + 0.012982106767594814, + 0.01093527115881443, + 0.07267389446496964, + 0.01982838287949562, + 0.005909726954996586, + 0.01079303678125143, + 0.01693347468972206, + 0.04506830498576164, + 0.004375000949949026, + 0.0051664444617927074, + 0.00549293402582407, + 0.003919671755284071, + 0.005825225263834, + 0.004940181039273739, + 0.008238508366048336, + 0.0044762552715837955, + 0.0031865115743130445, + 0.002439027652144432, + 0.006133535411208868, + 0.003927656915038824, + 0.003192062024027109, + 0.002578493906185031, + 0.0035046509001404047, + 0.007778163533657789, + 0.010195930488407612, + 0.01079606730490923, + 0.008174136281013489, + 0.014601089060306549, + 0.009685825556516647, + 0.007279772311449051, + 0.008361654356122017, + 0.011459087952971458, + 0.011410824954509735, + 0.016686936840415, + 0.01582728885114193, + 0.01615331508219242, + 0.024929381906986237, + 0.012313373386859894, + 0.014106732793152332, + 0.01319610234349966, + 0.005183023866266012, + 0.0036117674317210913, + 0.004692917689681053, + 0.007159537635743618, + 0.0021922322921454906, + 0.002727632410824299, + 0.0029073436744511127, + 0.002743585966527462, + 0.012538628652691841, + 0.006620374973863363, + 0.009934025816619396, + 0.006771286018192768, + 0.004344214219599962, + 0.005886984523385763, + 0.0062964120879769325, + 0.009117896668612957, + 0.01009275857359171, + 0.005093327257782221, + 0.007628198713064194, + 0.006587766110897064, + 0.003709618002176285, + 0.002746513346210122, + 0.0054218047298491, + 0.006680822465568781, + 0.01443017739802599, + 0.018200049176812172, + 0.022269194945693016, + 0.02008899115025997, + 0.008374154567718506, + 0.007805109955370426, + 0.01376363169401884, + 0.016280917450785637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.16478261351585388 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026365121826529503 + ], + "opposite_logprob_tokens": [ + -3.651365041732788 + ], + "logprob_reward": -0.026365121826529503, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026365121826529503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4537123441696167, + 0.1270364671945572, + 0.02872590161859989, + 0.041864000260829926, + 0.0452534444630146, + 0.13618019223213196, + 0.032271284610033035, + 0.034247178584337234, + 0.04858658090233803, + 0.052122537046670914 + ], + "credit_map": [ + 0.0325007364153862, + 0.04363928735256195, + 0.09756112098693848, + 0.05828819051384926, + 0.05081488564610481, + 0.057567041367292404, + 0.04527299478650093, + 0.06806809455156326, + 0.008057848550379276, + 0.008563993498682976, + 0.045612264424562454, + 0.01675545983016491, + 0.0053225327283144, + 0.0077407448552548885, + 0.010367926210165024, + 0.02461569756269455, + 0.003958555404096842, + 0.003847198560833931, + 0.00390642648562789, + 0.004982492886483669, + 0.0028134973254054785, + 0.0022170471493154764, + 0.004180437885224819, + 0.0028202447574585676, + 0.0032538585364818573, + 0.0026101742405444384, + 0.008897056803107262, + 0.003400494810193777, + 0.003800803329795599, + 0.0038772195111960173, + 0.0028151555452495813, + 0.013209237717092037, + 0.005052166059613228, + 0.006875831633806229, + 0.004550772253423929, + 0.008146108128130436, + 0.004958420060575008, + 0.004923053085803986, + 0.004858603235334158, + 0.005888488609343767, + 0.008466843515634537, + 0.014964616857469082, + 0.010784367099404335, + 0.014069363474845886, + 0.03367602080106735, + 0.014819157309830189, + 0.023337339982390404, + 0.016062501817941666, + 0.002343767089769244, + 0.0028943417128175497, + 0.0040502301417291164, + 0.008232859894633293, + 0.004263699986040592, + 0.003557773306965828, + 0.003279544645920396, + 0.0036490692291408777, + 0.005832088645547628, + 0.003399726003408432, + 0.004796481225639582, + 0.004459149669855833, + 0.0026484320405870676, + 0.004000066313892603, + 0.004727534018456936, + 0.0043837022967636585, + 0.009245161898434162, + 0.005562280770391226, + 0.008828056044876575, + 0.006456761155277491, + 0.0035898557398468256, + 0.0029270981904119253, + 0.005267825443297625, + 0.0067095402628183365, + 0.006877157371491194, + 0.008403493091464043, + 0.009758462198078632, + 0.007773207034915686, + 0.0048445891588926315, + 0.003450570162385702, + 0.004791283048689365, + 0.006223776377737522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.026365121826529503 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002648711204529 + ], + "opposite_logprob_tokens": [ + -2.695026397705078 + ], + "logprob_reward": -0.07002648711204529, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002648711204529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15815739333629608, + 0.16508391499519348, + 0.03619390353560448, + 0.3598777949810028, + 0.02964722365140915, + 0.14303742349147797, + 0.016208017244935036, + 0.026528457179665565, + 0.04710354283452034, + 0.018162311986088753 + ], + "credit_map": [ + 0.018994217738509178, + 0.013819221407175064, + 0.021793827414512634, + 0.01403147540986538, + 0.011751343496143818, + 0.026437774300575256, + 0.02020200341939926, + 0.03112752176821232, + 0.00979960523545742, + 0.008695043623447418, + 0.06243613362312317, + 0.018772713840007782, + 0.004001364577561617, + 0.008248064666986465, + 0.014585226774215698, + 0.038545772433280945, + 0.003040079725906253, + 0.0039945244789123535, + 0.005870831198990345, + 0.004610121250152588, + 0.0043583097867667675, + 0.003130112076178193, + 0.0071611423045396805, + 0.004028786439448595, + 0.022191859781742096, + 0.01875738985836506, + 0.09517025202512741, + 0.022760771214962006, + 0.012666752561926842, + 0.022050919011235237, + 0.013285499066114426, + 0.15299436450004578, + 0.0039917780086398125, + 0.004330754280090332, + 0.003157134633511305, + 0.004337255842983723, + 0.0034450895618647337, + 0.003156905062496662, + 0.0032413548324257135, + 0.003986951895058155, + 0.010177453979849815, + 0.016403285786509514, + 0.011613743379712105, + 0.013360104523599148, + 0.03533285856246948, + 0.015339178964495659, + 0.022764811292290688, + 0.018045982345938683, + 0.0016471513081341982, + 0.002011604141443968, + 0.0026042647659778595, + 0.0029150715563446283, + 0.001750171068124473, + 0.0015899245627224445, + 0.0018200913909822702, + 0.0018697394989430904, + 0.0043418207205832005, + 0.003434213111177087, + 0.005424157716333866, + 0.0034325518645346165, + 0.001846394850872457, + 0.002658180659636855, + 0.0027076106052845716, + 0.0026835284661501646, + 0.010066666640341282, + 0.005193214863538742, + 0.006056551355868578, + 0.007218244019895792, + 0.004021468106657267, + 0.0029988246969878674, + 0.004460151772946119, + 0.007088423240929842, + 0.0017964509315788746, + 0.0021375161595642567, + 0.003197216894477606, + 0.00273685809224844, + 0.0017929107416421175, + 0.0014269653474912047, + 0.002675506053492427, + 0.0023988871835172176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.07002648711204529 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904927968978882 + ], + "opposite_logprob_tokens": [ + -2.4640493392944336 + ], + "logprob_reward": -0.08904927968978882, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904927968978882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22097761929035187, + 0.24915607273578644, + 0.044174905866384506, + 0.13581259548664093, + 0.029989387840032578, + 0.11273513734340668, + 0.023806337267160416, + 0.02896878495812416, + 0.1303321123123169, + 0.024047067388892174 + ], + "credit_map": [ + 0.026281194761395454, + 0.02464948035776615, + 0.03926512971520424, + 0.027223685756325722, + 0.02754904516041279, + 0.028433986008167267, + 0.0208699032664299, + 0.026705175638198853, + 0.014432539232075214, + 0.013035189360380173, + 0.09424963593482971, + 0.02776295877993107, + 0.005421921145170927, + 0.013127670623362064, + 0.02156728506088257, + 0.05955887958407402, + 0.004022267181426287, + 0.004824410658329725, + 0.006788924802094698, + 0.008746370673179626, + 0.0061398153193295, + 0.0032538236118853092, + 0.005203587468713522, + 0.005195703823119402, + 0.009251915849745274, + 0.007387765683233738, + 0.032000258564949036, + 0.01178619172424078, + 0.007531533017754555, + 0.009092986583709717, + 0.0072203041054308414, + 0.05154163017868996, + 0.003037483897060156, + 0.003925483673810959, + 0.0033076631370931864, + 0.004586047027260065, + 0.003034858964383602, + 0.005099916830658913, + 0.00289508537389338, + 0.0041028461419045925, + 0.007823453284800053, + 0.013512462377548218, + 0.00882323645055294, + 0.011311964131891727, + 0.02739926241338253, + 0.012452381663024426, + 0.01744726113975048, + 0.01396510936319828, + 0.0020564543083310127, + 0.002381733851507306, + 0.0031644224654883146, + 0.004150465130805969, + 0.0034848530776798725, + 0.0018021485302597284, + 0.004038783255964518, + 0.0027274740859866142, + 0.005672667175531387, + 0.002932913601398468, + 0.0038761456962674856, + 0.0034656566567718983, + 0.0023106203880161047, + 0.0028188759461045265, + 0.0027105894405394793, + 0.0051813130266964436, + 0.024111630395054817, + 0.014436043798923492, + 0.02263285592198372, + 0.016070296987891197, + 0.010346192866563797, + 0.010348591022193432, + 0.013525822199881077, + 0.018860656768083572, + 0.002437035320326686, + 0.002269235672429204, + 0.0033373816404491663, + 0.0031302215065807104, + 0.004022474400699139, + 0.0032869388815015554, + 0.003228024346753955, + 0.0023357539903372526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.08904927968978882 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702937424182892 + ], + "opposite_logprob_tokens": [ + -2.1270294189453125 + ], + "logprob_reward": -0.12702937424182892, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702937424182892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14592796564102173, + 0.22927060723304749, + 0.13314881920814514, + 0.04558908939361572, + 0.07925461232662201, + 0.058493174612522125, + 0.09784674644470215, + 0.05792303383350372, + 0.09600040316581726, + 0.05654557794332504 + ], + "credit_map": [ + 0.023965151980519295, + 0.017087073996663094, + 0.017053525894880295, + 0.03104430064558983, + 0.012533373199403286, + 0.0111866295337677, + 0.016103550791740417, + 0.01695435307919979, + 0.013335604220628738, + 0.01726403646171093, + 0.07880541682243347, + 0.02481389418244362, + 0.00654069148004055, + 0.013200253248214722, + 0.020064041018486023, + 0.05524665117263794, + 0.013735729269683361, + 0.016546079888939857, + 0.01620505191385746, + 0.019787538796663284, + 0.01977299526333809, + 0.010086440481245518, + 0.022880572825670242, + 0.014134397730231285, + 0.0034004796762019396, + 0.002761976094916463, + 0.009777681902050972, + 0.0032242494635283947, + 0.001966034760698676, + 0.0043183485977351665, + 0.0040615014731884, + 0.016078811138868332, + 0.014017089270055294, + 0.009765524417161942, + 0.009535381570458412, + 0.016046084463596344, + 0.008122256956994534, + 0.00677903275936842, + 0.007016223389655352, + 0.007973010651767254, + 0.00791347585618496, + 0.006069582421332598, + 0.00797371007502079, + 0.00962731521576643, + 0.00636414997279644, + 0.005446084775030613, + 0.007420124486088753, + 0.007678717840462923, + 0.0117447879165411, + 0.009997033514082432, + 0.016579091548919678, + 0.021915581077337265, + 0.007433827500790358, + 0.009421498514711857, + 0.009542476385831833, + 0.011212438344955444, + 0.007738096173852682, + 0.005609128158539534, + 0.008013566955924034, + 0.007978534325957298, + 0.006198750343173742, + 0.008380393497645855, + 0.007349977735430002, + 0.006654578726738691, + 0.012145276181399822, + 0.011091116815805435, + 0.014567682519555092, + 0.013369172811508179, + 0.009288055822253227, + 0.01328950934112072, + 0.010777708142995834, + 0.011471872217953205, + 0.00695262523368001, + 0.006707795429974794, + 0.008123153820633888, + 0.007652204483747482, + 0.00545116513967514, + 0.004979012534022331, + 0.008357985876500607, + 0.00832163356244564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.12702937424182892 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a slightly high-angle medium close-up, the camera smoothly pans to the right, its movements perfectly synchronized with the meticulous movements of a seamstress working in her sunlit studio. The central subject is her hands, one holding a piece of emerald green silk flat on the surface of a worn, light oak cutting table, the other guiding a pair of heavy, gleaming chrome-plated tailor's scissors. Seen from the side, she is an elderly woman, her silver hair tied in a neat bun, her face a look of intense concentration. The scissors move steadily and deliberately, the polished steel blades opening and closing with a satisfyingly crisp rhythm along a faint chalk line, parting the uneven fabric with unwavering precision. The scene is illuminated by soft natural light streaming in from the large windows, creating a gentle sheen on the silk and highlighting the fine grain of the wooden table. In the softly blurred background, colorful fabric rolls, spools, and an old-fashioned sewing machine create a rich, textured backdrop. The atmosphere is one of quiet concentration and exquisite craftsmanship, a serene and intimate portrait of an artisan lost in her craft.", + "seed": 508, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_12/index_707_seed_508_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10407843515276909, + "vlm_logprob": -0.10407843515276909, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.10407843515276909, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22858355939388275 + ], + "opposite_logprob_tokens": [ + -1.603583574295044 + ], + "logprob_reward": -0.22858355939388275, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22858355939388275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15845830738544464, + 0.14899711310863495, + 0.13452653586864471, + 0.13433870673179626, + 0.17563998699188232, + 0.05583301931619644, + 0.03264204040169716, + 0.08354455232620239, + 0.03352818638086319, + 0.04249159246683121 + ], + "credit_map": [ + 0.040537185966968536, + 0.010342075489461422, + 0.043692704290151596, + 0.015252938494086266, + 0.01096498966217041, + 0.014719203114509583, + 0.01164020411670208, + 0.011308996938169003, + 0.02660481631755829, + 0.012042374350130558, + 0.05346662551164627, + 0.014014682732522488, + 0.010400700382888317, + 0.011426806449890137, + 0.011276066303253174, + 0.009765038266777992, + 0.00938896182924509, + 0.03926479071378708, + 0.007551948074251413, + 0.009023552760481834, + 0.009486661292612553, + 0.009016459807753563, + 0.013955415226519108, + 0.03683874383568764, + 0.013043784536421299, + 0.012034864164888859, + 0.009408082813024521, + 0.015017938800156116, + 0.01757579855620861, + 0.03035389445722103, + 0.02317316271364689, + 0.013731177896261215, + 0.013517994433641434, + 0.054149359464645386, + 0.010235622525215149, + 0.010941511951386929, + 0.012815970927476883, + 0.02829073928296566, + 0.03572279214859009, + 0.009965996257960796, + 0.007394640706479549, + 0.0076064905151724815, + 0.005974582862108946, + 0.006034787744283676, + 0.008249160833656788, + 0.008856487460434437, + 0.006561558693647385, + 0.005155312363058329, + 0.0053223357535898685, + 0.005005743820220232, + 0.003543796483427286, + 0.00379959843121469, + 0.004458093084394932, + 0.004038987215608358, + 0.003711296711117029, + 0.0027621882036328316, + 0.010469063185155392, + 0.012702982872724533, + 0.009591793641448021, + 0.012903912924230099, + 0.010990027338266373, + 0.009127829223871231, + 0.010183295235037804, + 0.007575646974146366, + 0.004179837182164192, + 0.0041412771679461, + 0.00436842255294323, + 0.004964219871908426, + 0.004379132762551308, + 0.004069861955940723, + 0.004663347266614437, + 0.002762089716270566, + 0.0048025199212133884, + 0.005633877590298653, + 0.0060625276528298855, + 0.005261346697807312, + 0.00812550075352192, + 0.00542880455031991, + 0.003108085598796606, + 0.004068928305059671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.22858355939388275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263676755130291 + ], + "opposite_logprob_tokens": [ + -3.651367664337158 + ], + "logprob_reward": -0.0263676755130291, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263676755130291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22144587337970734, + 0.15369929373264313, + 0.08432190865278244, + 0.08463113009929657, + 0.05858013033866882, + 0.060965124517679214, + 0.03845354914665222, + 0.21289095282554626, + 0.04264504835009575, + 0.042366959154605865 + ], + "credit_map": [ + 0.020107662305235863, + 0.010523534379899502, + 0.07946167141199112, + 0.01856306567788124, + 0.012681830674409866, + 0.05422050505876541, + 0.009523297660052776, + 0.016364319249987602, + 0.019646964967250824, + 0.01947365328669548, + 0.01555320993065834, + 0.013575366698205471, + 0.017555294558405876, + 0.02188831754028797, + 0.026666363701224327, + 0.019340135157108307, + 0.007456379476934671, + 0.02460046298801899, + 0.005686118267476559, + 0.007446083705872297, + 0.007958335801959038, + 0.0067606656812131405, + 0.009235377423465252, + 0.015178482048213482, + 0.01138838566839695, + 0.008245301432907581, + 0.012471846304833889, + 0.01439683511853218, + 0.007785331457853317, + 0.01030720490962267, + 0.009886945597827435, + 0.010149288922548294, + 0.004308042582124472, + 0.01658431626856327, + 0.005744076333940029, + 0.005488618277013302, + 0.004847368225455284, + 0.006468289066106081, + 0.011253893375396729, + 0.0038855266757309437, + 0.006121515296399593, + 0.010787642560899258, + 0.004837649408727884, + 0.010045276954770088, + 0.009145205840468407, + 0.008315330371260643, + 0.007751216180622578, + 0.003961282782256603, + 0.005644927266985178, + 0.004487429279834032, + 0.005677131004631519, + 0.003824663581326604, + 0.0042916857637465, + 0.004805202595889568, + 0.0060348366387188435, + 0.0036876762751489878, + 0.02680048905313015, + 0.03380211442708969, + 0.025577064603567123, + 0.0339825414121151, + 0.02424868941307068, + 0.024442942813038826, + 0.02501165308058262, + 0.019025452435016632, + 0.003883576486259699, + 0.005622074939310551, + 0.006075473967939615, + 0.005435765255242586, + 0.005455192178487778, + 0.00454761553555727, + 0.006251742132008076, + 0.0053736069239676, + 0.004237838089466095, + 0.003887265222147107, + 0.005526545457541943, + 0.005102122202515602, + 0.0066610341891646385, + 0.005702181253582239, + 0.005373975727707148, + 0.005875997710973024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0263676755130291 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043062712997198105 + ], + "opposite_logprob_tokens": [ + -3.168062686920166 + ], + "logprob_reward": -0.043062712997198105, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043062712997198105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27017638087272644, + 0.17021313309669495, + 0.15676742792129517, + 0.02798660285770893, + 0.04664643108844757, + 0.04666895791888237, + 0.0335937961935997, + 0.17523421347141266, + 0.03499017655849457, + 0.03772296383976936 + ], + "credit_map": [ + 0.07170078158378601, + 0.014170430600643158, + 0.09033558517694473, + 0.02644258178770542, + 0.016211576759815216, + 0.02415277250111103, + 0.011628379113972187, + 0.015534265898168087, + 0.025726202875375748, + 0.009256417863070965, + 0.06569703668355942, + 0.016751138493418694, + 0.009732861071825027, + 0.014902719296514988, + 0.01107469480484724, + 0.017072053626179695, + 0.015288633294403553, + 0.012562597170472145, + 0.01759510301053524, + 0.016431506723165512, + 0.01851465553045273, + 0.014553402550518513, + 0.024841846898198128, + 0.03697967529296875, + 0.0032208438497036695, + 0.002423830796033144, + 0.0036142461467534304, + 0.0031168689019978046, + 0.005086727440357208, + 0.003742926986888051, + 0.003452743636444211, + 0.003328414401039481, + 0.004220436792820692, + 0.00868956558406353, + 0.0051972041837871075, + 0.005054996348917484, + 0.0038451505824923515, + 0.005720837973058224, + 0.009329473599791527, + 0.00458876695483923, + 0.004782248288393021, + 0.00592945097014308, + 0.0036012944765388966, + 0.004501974675804377, + 0.01077037863433361, + 0.006766424980014563, + 0.005569915287196636, + 0.0047472682781517506, + 0.004133644048124552, + 0.004080032464116812, + 0.0042299535125494, + 0.0040029981173574924, + 0.0033962614834308624, + 0.005028546322137117, + 0.004237095359712839, + 0.0044852658174932, + 0.02233840525150299, + 0.028843170031905174, + 0.02071903459727764, + 0.02809055894613266, + 0.02027391642332077, + 0.0203519556671381, + 0.019444487988948822, + 0.015172694809734821, + 0.0044184960424900055, + 0.0033363911788910627, + 0.005602051969617605, + 0.004946841858327389, + 0.0051806410774588585, + 0.003474407596513629, + 0.0048422496765851974, + 0.003189097624272108, + 0.0049989656545221806, + 0.003983760252594948, + 0.007116422522813082, + 0.003979796078056097, + 0.005488013383001089, + 0.004020222928375006, + 0.004132556263357401, + 0.0040032281540334225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.043062712997198105 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062052879482507706 + ], + "opposite_logprob_tokens": [ + -2.8120529651641846 + ], + "logprob_reward": -0.062052879482507706, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062052879482507706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3326593339443207, + 0.12211111187934875, + 0.10065028816461563, + 0.040656473487615585, + 0.07328318059444427, + 0.029337698593735695, + 0.03262455016374588, + 0.2015899270772934, + 0.04198926314711571, + 0.02509814128279686 + ], + "credit_map": [ + 0.07227334380149841, + 0.024398816749453545, + 0.11273780465126038, + 0.035752687603235245, + 0.021135114133358, + 0.022796496748924255, + 0.0226480383425951, + 0.020917044952511787, + 0.014621896669268608, + 0.008069639094173908, + 0.04015364125370979, + 0.013313596136868, + 0.009001201950013638, + 0.013940300792455673, + 0.009660458192229271, + 0.013350378721952438, + 0.00992678478360176, + 0.018838630989193916, + 0.007725489791482687, + 0.00874821376055479, + 0.008854360319674015, + 0.008609438315033913, + 0.011815468780696392, + 0.026131901890039444, + 0.004701560828834772, + 0.004585710354149342, + 0.005174118094146252, + 0.008000973612070084, + 0.003957338631153107, + 0.006072875577956438, + 0.004319499246776104, + 0.003844397608190775, + 0.005031576380133629, + 0.022113030776381493, + 0.004286530427634716, + 0.004242392256855965, + 0.004922367632389069, + 0.007084159646183252, + 0.019327018409967422, + 0.006276103667914867, + 0.0029696663841605186, + 0.0035449713468551636, + 0.002686243737116456, + 0.004792466294020414, + 0.004314881283789873, + 0.00396368745714426, + 0.003393372753635049, + 0.003672409802675247, + 0.005949944257736206, + 0.003811601782217622, + 0.004170206841081381, + 0.004038631450384855, + 0.003287309082224965, + 0.0036649596877396107, + 0.004468015395104885, + 0.003233881900086999, + 0.024215294048190117, + 0.030739348381757736, + 0.023736968636512756, + 0.034205757081508636, + 0.022981911897659302, + 0.023013342171907425, + 0.024355312809348106, + 0.018341992050409317, + 0.005460020620375872, + 0.00594242662191391, + 0.005092254839837551, + 0.005550877191126347, + 0.006991025060415268, + 0.004860065411776304, + 0.004638496786355972, + 0.003454097080975771, + 0.0033768904395401478, + 0.003294119145721197, + 0.0038614885415881872, + 0.0028262536507099867, + 0.0030497743282467127, + 0.003104385221377015, + 0.002960052341222763, + 0.002625176915898919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.062052879482507706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032534837722778 + ], + "opposite_logprob_tokens": [ + -1.910325288772583 + ], + "logprob_reward": -0.16032534837722778, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032534837722778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40002718567848206, + 0.16291578114032745, + 0.22976236045360565, + 0.0799679085612297, + 0.051039740443229675, + 0.03078651614487171, + 0.011359107680618763, + 0.009193997830152512, + 0.017126260325312614, + 0.007821117527782917 + ], + "credit_map": [ + 0.11549640446901321, + 0.01980556733906269, + 0.08255879580974579, + 0.03437529131770134, + 0.022686272859573364, + 0.07893669605255127, + 0.02371923439204693, + 0.022448915988206863, + 0.028705842792987823, + 0.016668930649757385, + 0.06251993030309677, + 0.018171560019254684, + 0.008212566375732422, + 0.010622089728713036, + 0.009176187217235565, + 0.008838662877678871, + 0.014677383005619049, + 0.07393602281808853, + 0.00893568154424429, + 0.01171287801116705, + 0.016414638608694077, + 0.013692204840481281, + 0.019112898036837578, + 0.07128063589334488, + 0.010113963857293129, + 0.007261260412633419, + 0.00910462997853756, + 0.011385107412934303, + 0.010459981858730316, + 0.012176508083939552, + 0.008330835960805416, + 0.011135614477097988, + 0.004609508439898491, + 0.006723849102854729, + 0.002507525496184826, + 0.0038991193287074566, + 0.005318243056535721, + 0.00908011943101883, + 0.0069920457899570465, + 0.01190932746976614, + 0.002886482048779726, + 0.0037879932206124067, + 0.0019425603095442057, + 0.0044868504628539085, + 0.0045934575609862804, + 0.006072020623832941, + 0.002993183908984065, + 0.004023965913802385, + 0.0012937572319060564, + 0.0017013364704325795, + 0.000918416422791779, + 0.0013030453119426966, + 0.0010061266366392374, + 0.0018593029817566276, + 0.0015059212455525994, + 0.0017712005646899343, + 0.0010706321336328983, + 0.0011541941203176975, + 0.0014013689942657948, + 0.0011144609889015555, + 0.0009192505967803299, + 0.0011151129147037864, + 0.0011525428853929043, + 0.001266435137949884, + 0.0016464737709611654, + 0.0023379570338875055, + 0.0019061800558120012, + 0.0021752342581748962, + 0.0031599635258316994, + 0.0020579765550792217, + 0.002082541584968567, + 0.001759932260029018, + 0.0010821650503203273, + 0.0008165446342900395, + 0.0012732360046356916, + 0.0009079145966097713, + 0.0014245526399463415, + 0.0008758226758800447, + 0.0007606944418512285, + 0.0006801869021728635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.16032534837722778 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a slightly high-angle medium close-up, the camera smoothly pans to the right, its movements perfectly synchronized with the meticulous movements of a seamstress working in her sunlit studio. The central subject is her hands, one holding a piece of emerald green silk flat on the surface of a worn, light oak cutting table, the other guiding a pair of heavy, gleaming chrome-plated tailor's scissors. Seen from the side, she is an elderly woman, her silver hair tied in a neat bun, her face a look of intense concentration. The scissors move steadily and deliberately, the polished steel blades opening and closing with a satisfyingly crisp rhythm along a faint chalk line, parting the uneven fabric with unwavering precision. The scene is illuminated by soft natural light streaming in from the large windows, creating a gentle sheen on the silk and highlighting the fine grain of the wooden table. In the softly blurred background, colorful fabric rolls, spools, and an old-fashioned sewing machine create a rich, textured backdrop. The atmosphere is one of quiet concentration and exquisite craftsmanship, a serene and intimate portrait of an artisan lost in her craft.", + "seed": 509, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_12/index_707_seed_509_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1931595228612423, + "vlm_logprob": -0.1931595228612423, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.1931595228612423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6412121057510376 + ], + "opposite_logprob_tokens": [ + -0.7662121057510376 + ], + "logprob_reward": -0.6412121057510376, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6412121057510376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07643365114927292, + 0.08064980804920197, + 0.12222317606210709, + 0.03602862358093262, + 0.04748576506972313, + 0.5655480623245239, + 0.013724319636821747, + 0.01390804722905159, + 0.009368913248181343, + 0.03462965413928032 + ], + "credit_map": [ + 0.019865311682224274, + 0.008869829587638378, + 0.013563653454184532, + 0.01046114880591631, + 0.004856915213167667, + 0.004134942311793566, + 0.009677231311798096, + 0.0050046187825500965, + 0.02198568545281887, + 0.0049088685773313046, + 0.015791194513440132, + 0.010499674826860428, + 0.006244674324989319, + 0.0064600370824337006, + 0.00637411093339324, + 0.008385565131902695, + 0.016368452459573746, + 0.010476908646523952, + 0.02107389084994793, + 0.018190555274486542, + 0.010270455852150917, + 0.011434201151132584, + 0.01818516105413437, + 0.016223547980189323, + 0.0040819644927978516, + 0.004549421835690737, + 0.0053561050444841385, + 0.00797423254698515, + 0.003082343377172947, + 0.0040870364755392075, + 0.0033986081834882498, + 0.003498914185911417, + 0.016559554263949394, + 0.0029786089435219765, + 0.013377288356423378, + 0.004217588808387518, + 0.004359926097095013, + 0.0018356845248490572, + 0.0016231014160439372, + 0.002534018363803625, + 0.13838845491409302, + 0.09909802675247192, + 0.10680750757455826, + 0.11412305384874344, + 0.026522347703576088, + 0.0278259739279747, + 0.020827505737543106, + 0.031955182552337646, + 0.00208397232927382, + 0.001636723754927516, + 0.0024778596125543118, + 0.0013933825539425015, + 0.0007609797758050263, + 0.0010283378651365638, + 0.0012096223654225469, + 0.0031334415543824434, + 0.0017963587306439877, + 0.0016639980021864176, + 0.0018262609373778105, + 0.0014926516450941563, + 0.002621134277433157, + 0.0015275973128154874, + 0.0017794050509110093, + 0.0012006411561742425, + 0.0019670985639095306, + 0.0015005550812929869, + 0.0018083108589053154, + 0.0014492865884676576, + 0.000428383908001706, + 0.0005586143815889955, + 0.0007106601842679083, + 0.0009460038272663951, + 0.0070087602362036705, + 0.002297711092978716, + 0.0064854696393013, + 0.003985266666859388, + 0.0023240901064127684, + 0.002999765332788229, + 0.004396315198391676, + 0.00513227516785264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.6412121057510376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020616523921489716 + ], + "opposite_logprob_tokens": [ + -3.8956165313720703 + ], + "logprob_reward": -0.020616523921489716, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020616523921489716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1778206080198288, + 0.10646384209394455, + 0.2922230660915375, + 0.0644235610961914, + 0.04327613487839699, + 0.12313317507505417, + 0.02927975542843342, + 0.042480744421482086, + 0.035535406321287155, + 0.08536364138126373 + ], + "credit_map": [ + 0.027451172471046448, + 0.013486488722264767, + 0.02710733562707901, + 0.023841267451643944, + 0.020210647955536842, + 0.032383594661951065, + 0.02130805142223835, + 0.012032055296003819, + 0.010856314562261105, + 0.00941278412938118, + 0.021814830601215363, + 0.015124416910111904, + 0.012679356150329113, + 0.009895515628159046, + 0.015784481540322304, + 0.010896146297454834, + 0.02992362529039383, + 0.021703358739614487, + 0.05407828092575073, + 0.03909394517540932, + 0.024170486256480217, + 0.02776270918548107, + 0.050516240298748016, + 0.044974446296691895, + 0.006033480633050203, + 0.007440243382006884, + 0.0097119752317667, + 0.010759755969047546, + 0.004829501733183861, + 0.006081158760935068, + 0.009389448910951614, + 0.01017800159752369, + 0.00674622505903244, + 0.004297568928450346, + 0.01055630762130022, + 0.005537796299904585, + 0.005393615458160639, + 0.0031196598429232836, + 0.003645051270723343, + 0.003979911096394062, + 0.02753281407058239, + 0.018702201545238495, + 0.02255902998149395, + 0.025413116440176964, + 0.007631457410752773, + 0.009267057292163372, + 0.0054413373582065105, + 0.006586161442101002, + 0.0041083283722400665, + 0.003260722616687417, + 0.0033821004908531904, + 0.003670639358460903, + 0.0023251520469784737, + 0.00252320128493011, + 0.0041558584198355675, + 0.005853752139955759, + 0.004091538023203611, + 0.004809834528714418, + 0.004791515879333019, + 0.004197404254227877, + 0.008862816728651524, + 0.004192985128611326, + 0.006507298909127712, + 0.005027352832257748, + 0.005502218380570412, + 0.0034139587078243494, + 0.007077058777213097, + 0.004592570010572672, + 0.0022538721095770597, + 0.0030598887242376804, + 0.00468353321775794, + 0.004952310584485531, + 0.01293959841132164, + 0.006873472593724728, + 0.013395749032497406, + 0.00871419720351696, + 0.0066365129314363, + 0.007480998057872057, + 0.014788078144192696, + 0.01453503780066967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.020616523921489716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062052879482507706 + ], + "opposite_logprob_tokens": [ + -2.8120529651641846 + ], + "logprob_reward": -0.062052879482507706, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062052879482507706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19334906339645386, + 0.06932501494884491, + 0.051425136625766754, + 0.07891437411308289, + 0.08041508495807648, + 0.33694201707839966, + 0.03951844200491905, + 0.02738899365067482, + 0.02973923645913601, + 0.09298259019851685 + ], + "credit_map": [ + 0.02704661898314953, + 0.015420052222907543, + 0.034550491720438004, + 0.025080285966396332, + 0.021155700087547302, + 0.03394077345728874, + 0.022534001618623734, + 0.013621124438941479, + 0.012357993051409721, + 0.005635784938931465, + 0.01575276255607605, + 0.009715858846902847, + 0.00671066390350461, + 0.005243566818535328, + 0.00813103374093771, + 0.005777354817837477, + 0.004794265143573284, + 0.004452927969396114, + 0.00814606249332428, + 0.008512654341757298, + 0.005034876987338066, + 0.00586290517821908, + 0.008859611116349697, + 0.005761829670518637, + 0.006602785550057888, + 0.009974123910069466, + 0.01276345830410719, + 0.014500772580504417, + 0.0046802908182144165, + 0.006746916100382805, + 0.01302080973982811, + 0.010625213384628296, + 0.02502104826271534, + 0.0052528418600559235, + 0.025211596861481667, + 0.007032700348645449, + 0.006724224891513586, + 0.004154039081186056, + 0.0030734564643353224, + 0.003945180214941502, + 0.08338204771280289, + 0.05817229673266411, + 0.06483802199363708, + 0.06355965882539749, + 0.016735943034291267, + 0.017453113570809364, + 0.012375886552035809, + 0.02042505331337452, + 0.005033839028328657, + 0.0028053291607648134, + 0.005804325919598341, + 0.005201990716159344, + 0.0037781596183776855, + 0.0035046313423663378, + 0.00639047846198082, + 0.006999687757343054, + 0.003200952894985676, + 0.0031019083689898252, + 0.00397680327296257, + 0.0025661103427410126, + 0.0036304874811321497, + 0.004425376188009977, + 0.003899352392181754, + 0.002588002011179924, + 0.004527014680206776, + 0.004363021347671747, + 0.0054906816221773624, + 0.0034375074319541454, + 0.0023601532448083162, + 0.0017190510407090187, + 0.003031886415556073, + 0.004809921141713858, + 0.018445540219545364, + 0.007053202483803034, + 0.010455830954015255, + 0.01311418041586876, + 0.006091698072850704, + 0.009952818974852562, + 0.009524090215563774, + 0.018345221877098083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.062052879482507706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062054336071014404 + ], + "opposite_logprob_tokens": [ + -2.812054395675659 + ], + "logprob_reward": -0.062054336071014404, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062054336071014404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15161973237991333, + 0.12627790868282318, + 0.2508164048194885, + 0.04372323304414749, + 0.06830278038978577, + 0.1367184817790985, + 0.03859087824821472, + 0.05492643639445305, + 0.03448289632797241, + 0.09454135596752167 + ], + "credit_map": [ + 0.019934864714741707, + 0.013640898279845715, + 0.030261315405368805, + 0.021298235282301903, + 0.017455270513892174, + 0.013074284419417381, + 0.02140198089182377, + 0.01455286517739296, + 0.030643269419670105, + 0.008409391157329082, + 0.02976486086845398, + 0.01914995349943638, + 0.009508898481726646, + 0.007026921957731247, + 0.010525437071919441, + 0.01124916598200798, + 0.022578278556466103, + 0.016774091869592667, + 0.051401447504758835, + 0.0350520983338356, + 0.020306041464209557, + 0.020428914576768875, + 0.04512457549571991, + 0.039150919765233994, + 0.006025563459843397, + 0.005816434510052204, + 0.005796846468001604, + 0.00954978633671999, + 0.0033494685776531696, + 0.004735891707241535, + 0.005011061672121286, + 0.0034381751902401447, + 0.020330118015408516, + 0.005399651359766722, + 0.0193344596773386, + 0.006950074806809425, + 0.0033378528896719217, + 0.00358527316711843, + 0.005016148556023836, + 0.004349194001406431, + 0.033329643309116364, + 0.01962064392864704, + 0.02619830332696438, + 0.02411474846303463, + 0.006638853810727596, + 0.008303099311888218, + 0.007678451482206583, + 0.010834715329110622, + 0.004645084962248802, + 0.0027031567879021168, + 0.007715994957834482, + 0.004604895133525133, + 0.0039008292369544506, + 0.0031978075858205557, + 0.004629763774573803, + 0.007193342316895723, + 0.006465581711381674, + 0.004888024181127548, + 0.008030517026782036, + 0.0052345613949000835, + 0.004842204041779041, + 0.006472615525126457, + 0.011194960214197636, + 0.007797963451594114, + 0.004802710842341185, + 0.0065668257884681225, + 0.005467391572892666, + 0.005474087782204151, + 0.002052511554211378, + 0.0031867017969489098, + 0.002210579114034772, + 0.004722083453088999, + 0.01791728101670742, + 0.005660067778080702, + 0.018268072977662086, + 0.01078768540173769, + 0.005528733134269714, + 0.008853868581354618, + 0.01334719080477953, + 0.014178448356688023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.062054336071014404 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject an elderly woman with silver hair tied in a neat bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman working with a piece of emerald green silk on a worn, light oak cutting table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman using heavy, gleaming chrome-plated tailor's scissors to cut the fabric along a faint chalk line?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by soft natural light from large windows, creating a sheen on the silk and highlighting the wood grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly pan to the right while maintaining a slightly high-angle medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986176908016205 + ], + "opposite_logprob_tokens": [ + -1.8048617839813232 + ], + "logprob_reward": -0.17986176908016205, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986176908016205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06403607875108719, + 0.059000223875045776, + 0.04379074648022652, + 0.017984585836529732, + 0.014924467541277409, + 0.06853780150413513, + 0.020108336582779884, + 0.028855472803115845, + 0.021058518439531326, + 0.6617037653923035 + ], + "credit_map": [ + 0.008500169962644577, + 0.00486957561224699, + 0.00917520746588707, + 0.007751405239105225, + 0.008123471401631832, + 0.013410092331469059, + 0.006643571425229311, + 0.0055625964887440205, + 0.01616477034986019, + 0.0036201702896505594, + 0.012008141726255417, + 0.008366175927221775, + 0.004101913888007402, + 0.005634830333292484, + 0.004328817129135132, + 0.0047754161059856415, + 0.00489632086828351, + 0.004153012298047543, + 0.007744955830276012, + 0.006196380127221346, + 0.004431447479873896, + 0.005333725828677416, + 0.005635289940983057, + 0.005399617832154036, + 0.002040064660832286, + 0.0021426165476441383, + 0.0023681519087404013, + 0.004114517942070961, + 0.0013493498554453254, + 0.0014384245732799172, + 0.0019344916800037026, + 0.0025969697162508965, + 0.001859776908531785, + 0.0015023858286440372, + 0.0031017803121358156, + 0.002315347781404853, + 0.0012428024783730507, + 0.001206884509883821, + 0.0021065135952085257, + 0.0015889783389866352, + 0.014028646983206272, + 0.00802661944180727, + 0.012988075613975525, + 0.011578470468521118, + 0.006178798153996468, + 0.005624345038086176, + 0.004805763252079487, + 0.005307094659656286, + 0.0023574973456561565, + 0.0015034960815683007, + 0.004655345808714628, + 0.002676895819604397, + 0.001370918471366167, + 0.0017295876750722528, + 0.0020626087207347155, + 0.0037519873585551977, + 0.003872375236824155, + 0.0045971861109137535, + 0.003512541064992547, + 0.0023578996770083904, + 0.004883015062659979, + 0.00424233777448535, + 0.003171930555254221, + 0.0022181898821145296, + 0.00406164163723588, + 0.004032175987958908, + 0.002848264528438449, + 0.0022474315483123064, + 0.0019394769333302975, + 0.0018255648901686072, + 0.0018001324497163296, + 0.002303833607584238, + 0.11576414108276367, + 0.05974922701716423, + 0.07035830616950989, + 0.0774831473827362, + 0.04467414692044258, + 0.06490558385848999, + 0.06934262067079544, + 0.15942659974098206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.17986176908016205 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_13/index_tensor([365, 365])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_13/index_tensor([365, 365])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a3c3ba128084666f4f67c9e6ba71d0172ef3644e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_13/index_tensor([365, 365])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a steady tracking shot from a slightly upward angle, capturing a white egret soaring over a vast expanse of lush reeds swaying gently in the breeze. The egret's slender figure, its flawless snow-white feathers appearing remarkably pure in the soft, diffused morning light. Its broad wings flap gracefully, its slender black legs tucked back, its sharp yellow beak pointed forward, its movements fluid and deliberate. The world is enveloped in a hazy, milky morning mist. A veil-like haze rises from the tranquil surface of the lake, blurring distant scenery and creating an ethereal, tranquil atmosphere. The camera then steadily and smoothly pulls back, gradually reducing the egret's size to a nimble white dot. Its flight path traces a graceful arc across the vast canvas of golden reeds and pale gray sky, revealing a serene and majestic panorama of the wetland.", + "seed": 18375, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_13/index_365_seed_18375_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02723110429942608, + "vlm_logprob": -0.02723110429942608, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02723110429942608, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008154788054525852 + ], + "opposite_logprob_tokens": [ + -5.13315486907959 + ], + "logprob_reward": -0.008154788054525852, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008154788054525852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16445915400981903, + 0.25263941287994385, + 0.07410857826471329, + 0.040514931082725525, + 0.03177820146083832, + 0.08018634468317032, + 0.07909273356199265, + 0.13601481914520264, + 0.05361701548099518, + 0.08758878707885742 + ], + "credit_map": [ + 0.01846890337765217, + 0.014838146977126598, + 0.02873070538043976, + 0.02525140531361103, + 0.01890803501009941, + 0.018833642825484276, + 0.022185837849974632, + 0.01724250055849552, + 0.017235660925507545, + 0.014402748085558414, + 0.06784091144800186, + 0.024567442014813423, + 0.013635286130011082, + 0.029064329341053963, + 0.026362191885709763, + 0.05953085049986839, + 0.009639465250074863, + 0.008622956462204456, + 0.009765022434294224, + 0.009452557191252708, + 0.008743241429328918, + 0.010004064999520779, + 0.007239494938403368, + 0.010641777887940407, + 0.0039127725176513195, + 0.0043755145743489265, + 0.005153063219040632, + 0.004785500932484865, + 0.0068153319880366325, + 0.004194706212729216, + 0.007695593871176243, + 0.003582451492547989, + 0.0041319578886032104, + 0.003584717633202672, + 0.0040530795231461525, + 0.005677178967744112, + 0.003704679664224386, + 0.0043991026468575, + 0.003460703883320093, + 0.0027667840477079153, + 0.00824711099267006, + 0.00799105316400528, + 0.009623288176953793, + 0.010727808810770512, + 0.0076880669221282005, + 0.00921026524156332, + 0.017802704125642776, + 0.008896051906049252, + 0.0065811374224722385, + 0.009857680648565292, + 0.00889481883496046, + 0.01053414773195982, + 0.010920314118266106, + 0.01929212175309658, + 0.006957636680454016, + 0.00605488196015358, + 0.011901259422302246, + 0.034620512276887894, + 0.0105079784989357, + 0.006855752319097519, + 0.02903136797249317, + 0.015746818855404854, + 0.00945240631699562, + 0.01789872720837593, + 0.0035677612759172916, + 0.013111032545566559, + 0.004728315863758326, + 0.0043829637579619884, + 0.013941104523837566, + 0.005814012140035629, + 0.005132589023560286, + 0.0029392410069704056, + 0.007814855314791203, + 0.012622117064893246, + 0.006753746420145035, + 0.013452772051095963, + 0.012163735926151276, + 0.00928971916437149, + 0.0176590196788311, + 0.00783282145857811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.008154788054525852 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083466000854969 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.011083466000854969, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083466000854969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15800921618938446, + 0.2600696384906769, + 0.04805399104952812, + 0.0562577024102211, + 0.02428998425602913, + 0.08152489364147186, + 0.09171687811613083, + 0.10699959099292755, + 0.10072186589241028, + 0.07235618680715561 + ], + "credit_map": [ + 0.015581739135086536, + 0.013932296074926853, + 0.03817753866314888, + 0.017106391489505768, + 0.015457460656762123, + 0.02145162783563137, + 0.0173291377723217, + 0.018973030149936676, + 0.01727657951414585, + 0.01830407604575157, + 0.07108038663864136, + 0.026567615568637848, + 0.016890497878193855, + 0.030589548870921135, + 0.026759015396237373, + 0.05260192230343819, + 0.006784344557672739, + 0.00633691530674696, + 0.0065837823785841465, + 0.004343912936747074, + 0.004488445818424225, + 0.008040730841457844, + 0.004872794728726149, + 0.006603068672120571, + 0.0077575198374688625, + 0.01547045074403286, + 0.00666428729891777, + 0.005310067441314459, + 0.003850410459563136, + 0.007070573512464762, + 0.005436466541141272, + 0.004697930533438921, + 0.002791921142488718, + 0.002551011275500059, + 0.003478922415524721, + 0.003935059066861868, + 0.0028551171999424696, + 0.0031544025987386703, + 0.002963118953630328, + 0.0025604343973100185, + 0.009774471633136272, + 0.00833889376372099, + 0.01656060293316841, + 0.014528516680002213, + 0.005356048233807087, + 0.0071266693994402885, + 0.01056646928191185, + 0.009273224510252476, + 0.00782173965126276, + 0.01022371556609869, + 0.014756345190107822, + 0.012477659620344639, + 0.013047649525105953, + 0.013456443324685097, + 0.011614272370934486, + 0.008319061249494553, + 0.00966098252683878, + 0.017614973708987236, + 0.013509135693311691, + 0.01069494429975748, + 0.016986260190606117, + 0.009749229066073895, + 0.018058018758893013, + 0.01072605699300766, + 0.006339058745652437, + 0.028738945722579956, + 0.005889960564672947, + 0.00711101433262229, + 0.03118065930902958, + 0.008555477485060692, + 0.008625647984445095, + 0.004281116183847189, + 0.006507080513983965, + 0.011159923858940601, + 0.004748405423015356, + 0.00570583576336503, + 0.014079543761909008, + 0.00970908347517252, + 0.012029564008116722, + 0.008416757918894291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.011083466000854969 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594919100403786 + ], + "opposite_logprob_tokens": [ + -3.895594835281372 + ], + "logprob_reward": -0.020594919100403786, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594919100403786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16249221563339233, + 0.16575028002262115, + 0.06436289846897125, + 0.0748455822467804, + 0.038835298269987106, + 0.0920020267367363, + 0.029931604862213135, + 0.21408647298812866, + 0.10381615906953812, + 0.053877465426921844 + ], + "credit_map": [ + 0.017538024112582207, + 0.017315004020929337, + 0.026344917714595795, + 0.02022642455995083, + 0.021540336310863495, + 0.018397288396954536, + 0.024580946192145348, + 0.01654926873743534, + 0.012876646593213081, + 0.012272197753190994, + 0.038116924464702606, + 0.018583400174975395, + 0.013603754341602325, + 0.019540535286068916, + 0.020338673144578934, + 0.030418148264288902, + 0.006022977642714977, + 0.007655088324099779, + 0.0077969045378267765, + 0.007421051152050495, + 0.011048523709177971, + 0.008137687109410763, + 0.009359652176499367, + 0.006921017076820135, + 0.006653030402958393, + 0.009431455284357071, + 0.01152326725423336, + 0.008787761442363262, + 0.012080250307917595, + 0.008515680208802223, + 0.011617681011557579, + 0.006236451677978039, + 0.003393127117305994, + 0.0037059420719742775, + 0.004506049212068319, + 0.007412910461425781, + 0.0054871984757483006, + 0.005831225775182247, + 0.005147659219801426, + 0.003351187100633979, + 0.0094448858872056, + 0.008982065133750439, + 0.010649365372955799, + 0.010723358951508999, + 0.009574229829013348, + 0.012114929035305977, + 0.01819869503378868, + 0.01231449842453003, + 0.002563643967732787, + 0.0036481451243162155, + 0.003410802921280265, + 0.004370293579995632, + 0.004195484798401594, + 0.006059639621526003, + 0.003364624921232462, + 0.002318968763574958, + 0.014981422573328018, + 0.05634918808937073, + 0.016812004148960114, + 0.015859752893447876, + 0.050989408046007156, + 0.024596014991402626, + 0.014539000578224659, + 0.019959690049290657, + 0.006257867440581322, + 0.026282532140612602, + 0.007705125492066145, + 0.008021533489227295, + 0.028341224417090416, + 0.012741434387862682, + 0.007365910802036524, + 0.007100528106093407, + 0.008942641317844391, + 0.007730563636869192, + 0.0045002903789281845, + 0.004297293722629547, + 0.006835598964244127, + 0.006737926509231329, + 0.00938979722559452, + 0.005443357862532139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.020594919100403786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026333538815379143 + ], + "opposite_logprob_tokens": [ + -3.6513335704803467 + ], + "logprob_reward": -0.026333538815379143, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026333538815379143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16726329922676086, + 0.24181930720806122, + 0.10914558917284012, + 0.03824440389871597, + 0.034193575382232666, + 0.11248167604207993, + 0.05177107825875282, + 0.05114037171006203, + 0.14535567164421082, + 0.04858506843447685 + ], + "credit_map": [ + 0.016567537561058998, + 0.014682517386972904, + 0.03595750778913498, + 0.02252996526658535, + 0.015654323622584343, + 0.018995335325598717, + 0.02247508428990841, + 0.020401030778884888, + 0.016189323738217354, + 0.015039009042084217, + 0.0654660016298294, + 0.0220852829515934, + 0.015423603355884552, + 0.025822311639785767, + 0.025353262200951576, + 0.05644049867987633, + 0.01620239019393921, + 0.015284494496881962, + 0.02116227336227894, + 0.012476014904677868, + 0.007117103319615126, + 0.014754727482795715, + 0.00949324294924736, + 0.01265534944832325, + 0.004103881306946278, + 0.006249384954571724, + 0.004822102840989828, + 0.0038287220522761345, + 0.004733007401227951, + 0.004861454013735056, + 0.0067442250438034534, + 0.0029016255866736174, + 0.0037831771187484264, + 0.0035254922695457935, + 0.00544072687625885, + 0.005374222528189421, + 0.00441512418910861, + 0.003139364765956998, + 0.004168536979705095, + 0.004346929956227541, + 0.011641878634691238, + 0.011465191841125488, + 0.019463583827018738, + 0.020881138741970062, + 0.00616822624579072, + 0.0065001011826097965, + 0.01876075379550457, + 0.017600802704691887, + 0.0041669877246022224, + 0.005998620297759771, + 0.010026387870311737, + 0.008194662630558014, + 0.0068105063401162624, + 0.005536227021366358, + 0.006986820138990879, + 0.004050868097692728, + 0.003905895398929715, + 0.00605934439226985, + 0.008648233488202095, + 0.006199230439960957, + 0.006052413489669561, + 0.006419865880161524, + 0.009068429470062256, + 0.004786956124007702, + 0.013059251010417938, + 0.025256512686610222, + 0.021491223946213722, + 0.019130509346723557, + 0.0206941906362772, + 0.022021526470780373, + 0.009411945939064026, + 0.014290528371930122, + 0.004632860422134399, + 0.005624088924378157, + 0.0038846521638333797, + 0.005875586066395044, + 0.004895072430372238, + 0.0070167407393455505, + 0.009856411255896091, + 0.006799656432121992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.026333538815379143 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06998880952596664 + ], + "opposite_logprob_tokens": [ + -2.69498872756958 + ], + "logprob_reward": -0.06998880952596664, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06998880952596664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08799736201763153, + 0.1391829252243042, + 0.04297107830643654, + 0.04730948805809021, + 0.05299770087003708, + 0.1309666782617569, + 0.21761895716190338, + 0.06187104061245918, + 0.12848618626594543, + 0.09059863537549973 + ], + "credit_map": [ + 0.010628695599734783, + 0.009036842733621597, + 0.01138613373041153, + 0.015053872019052505, + 0.012842288240790367, + 0.011327666230499744, + 0.011620706878602505, + 0.006101150531321764, + 0.010244529694318771, + 0.009628881700336933, + 0.03659788519144058, + 0.013724028132855892, + 0.008967502042651176, + 0.015670251101255417, + 0.013930821791291237, + 0.03041902557015419, + 0.006239231675863266, + 0.005640345625579357, + 0.008347884751856327, + 0.005345330107957125, + 0.0036092689260840416, + 0.005361736752092838, + 0.00386791885830462, + 0.004559363704174757, + 0.004147848580032587, + 0.006274869665503502, + 0.006696547381579876, + 0.004966748412698507, + 0.007720344699919224, + 0.006678933277726173, + 0.006417019758373499, + 0.00440717488527298, + 0.005236118100583553, + 0.0058058020658791065, + 0.00676374975591898, + 0.007102012634277344, + 0.0074738492257893085, + 0.008631144650280476, + 0.006314499769359827, + 0.005670526996254921, + 0.010145435109734535, + 0.014256250113248825, + 0.018380669876933098, + 0.018477249890565872, + 0.009907045401632786, + 0.012494984082877636, + 0.028176147490739822, + 0.019128898158669472, + 0.01927240937948227, + 0.03166445344686508, + 0.025940900668501854, + 0.02955416962504387, + 0.03397766128182411, + 0.04143118858337402, + 0.018272558227181435, + 0.01750561222434044, + 0.006121412385255098, + 0.008636490441858768, + 0.008304641582071781, + 0.007756419945508242, + 0.00801586639136076, + 0.008556530810892582, + 0.008707803674042225, + 0.005771874450147152, + 0.00821718294173479, + 0.027572687715291977, + 0.010990777984261513, + 0.011457471176981926, + 0.03079311177134514, + 0.016014698892831802, + 0.013046438805758953, + 0.010393820703029633, + 0.008221861906349659, + 0.009733626618981361, + 0.006684602703899145, + 0.01896379142999649, + 0.011233695782721043, + 0.010066810995340347, + 0.017560629174113274, + 0.00813361257314682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.06998880952596664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a steady tracking shot from a slightly upward angle, capturing a white egret soaring over a vast expanse of lush reeds swaying gently in the breeze. The egret's slender figure, its flawless snow-white feathers appearing remarkably pure in the soft, diffused morning light. Its broad wings flap gracefully, its slender black legs tucked back, its sharp yellow beak pointed forward, its movements fluid and deliberate. The world is enveloped in a hazy, milky morning mist. A veil-like haze rises from the tranquil surface of the lake, blurring distant scenery and creating an ethereal, tranquil atmosphere. The camera then steadily and smoothly pulls back, gradually reducing the egret's size to a nimble white dot. Its flight path traces a graceful arc across the vast canvas of golden reeds and pale gray sky, revealing a serene and majestic panorama of the wetland.", + "seed": 18376, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_13/index_365_seed_18376_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03008935060352087, + "vlm_logprob": -0.03008935060352087, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03008935060352087, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015184991993010044 + ], + "opposite_logprob_tokens": [ + -4.5151848793029785 + ], + "logprob_reward": -0.015184991993010044, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015184991993010044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16046474874019623, + 0.053729746490716934, + 0.2358863800764084, + 0.020721009001135826, + 0.014406539499759674, + 0.029913214966654778, + 0.06976617127656937, + 0.07279075682163239, + 0.24806924164295197, + 0.09425216913223267 + ], + "credit_map": [ + 0.016436895355582237, + 0.01499397773295641, + 0.014131090603768826, + 0.008865025825798512, + 0.030826572328805923, + 0.040664706379175186, + 0.013983620330691338, + 0.020562872290611267, + 0.0076998695731163025, + 0.004687154665589333, + 0.011011544615030289, + 0.004473591223359108, + 0.005912399385124445, + 0.006845257710665464, + 0.006276806350797415, + 0.006823126692324877, + 0.024132762104272842, + 0.010767098516225815, + 0.02804730460047722, + 0.015720529481768608, + 0.015894779935479164, + 0.07196567952632904, + 0.016190145164728165, + 0.053168099373579025, + 0.0027657083701342344, + 0.0020028003491461277, + 0.0025204767007380724, + 0.0030410524923354387, + 0.0027721398510038853, + 0.0022632356267422438, + 0.0024978092405945063, + 0.002857789397239685, + 0.0017136229434981942, + 0.0015227706171572208, + 0.002129089552909136, + 0.0017534376820549369, + 0.0018494731048122048, + 0.0019168545259162784, + 0.0019413988338783383, + 0.0015798923559486866, + 0.004598267376422882, + 0.0027267266996204853, + 0.003405775176361203, + 0.0032791239209473133, + 0.00328036374412477, + 0.0044953785836696625, + 0.004507743753492832, + 0.003619837574660778, + 0.009138388559222221, + 0.0129401795566082, + 0.00684745516628027, + 0.006975564174354076, + 0.011784159578382969, + 0.00942879356443882, + 0.005200512707233429, + 0.007451125420629978, + 0.007506994530558586, + 0.00748481135815382, + 0.00819310825318098, + 0.012549378909170628, + 0.010154923424124718, + 0.007916699163615704, + 0.009474914520978928, + 0.00950993038713932, + 0.019107531756162643, + 0.006398150697350502, + 0.08216290920972824, + 0.03359108045697212, + 0.012152735143899918, + 0.05979530140757561, + 0.022314036265015602, + 0.012547518126666546, + 0.008722474798560143, + 0.010755036026239395, + 0.016669003292918205, + 0.011203190311789513, + 0.008512456901371479, + 0.02074793539941311, + 0.008439022116363049, + 0.009203054942190647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.015184991993010044 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020593751221895218 + ], + "opposite_logprob_tokens": [ + -3.8955936431884766 + ], + "logprob_reward": -0.020593751221895218, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020593751221895218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20921966433525085, + 0.050061240792274475, + 0.12065717577934265, + 0.04535367339849472, + 0.028251197189092636, + 0.04036435857415199, + 0.0791054219007492, + 0.13260382413864136, + 0.08909949660301208, + 0.2052839696407318 + ], + "credit_map": [ + 0.018787480890750885, + 0.013265139423310757, + 0.011170431040227413, + 0.008649787865579128, + 0.0407937727868557, + 0.0667974129319191, + 0.014832830987870693, + 0.03492279723286629, + 0.006132905371487141, + 0.0033882155548781157, + 0.01134230475872755, + 0.003626900725066662, + 0.00495660537853837, + 0.008595365099608898, + 0.005118320230394602, + 0.006900628563016653, + 0.011829945258796215, + 0.006213502027094364, + 0.01454822439700365, + 0.008975419215857983, + 0.008314022794365883, + 0.03379347547888756, + 0.010912230238318443, + 0.026070352643728256, + 0.004489727783948183, + 0.0031949959229677916, + 0.009636356495320797, + 0.005386460106819868, + 0.003983302041888237, + 0.00715635484084487, + 0.004315330181270838, + 0.0071911499835550785, + 0.0024473113007843494, + 0.002601195592433214, + 0.0035606648307293653, + 0.004879116080701351, + 0.00245233834721148, + 0.0030985840130597353, + 0.0050351242534816265, + 0.004176860675215721, + 0.004118040669709444, + 0.005266697145998478, + 0.004693520255386829, + 0.005289128981530666, + 0.004894919227808714, + 0.004533418454229832, + 0.006114805582910776, + 0.005453828256577253, + 0.009815704077482224, + 0.009883863851428032, + 0.011420723982155323, + 0.008024928160011768, + 0.005901163909584284, + 0.010521944612264633, + 0.010353500954806805, + 0.013183590024709702, + 0.012449613772332668, + 0.012759745121002197, + 0.015305157750844955, + 0.021393295377492905, + 0.022321492433547974, + 0.01599865034222603, + 0.014924880117177963, + 0.017450986430048943, + 0.014336216263473034, + 0.008059736341238022, + 0.015871575102210045, + 0.01249030139297247, + 0.008510654792189598, + 0.010231014341115952, + 0.008946005254983902, + 0.010653987526893616, + 0.015306766144931316, + 0.012101496569812298, + 0.05262316018342972, + 0.016892580315470695, + 0.018787797540426254, + 0.06071750819683075, + 0.016571782529354095, + 0.012282896786928177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.020593751221895218 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083701625466347 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.011083701625466347, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083701625466347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23214952647686005, + 0.06252087652683258, + 0.06210862845182419, + 0.02308773621916771, + 0.01644035056233406, + 0.030858248472213745, + 0.04099980741739273, + 0.05014246329665184, + 0.4100838005542755, + 0.07160865515470505 + ], + "credit_map": [ + 0.025133587419986725, + 0.028081508353352547, + 0.029941949993371964, + 0.017500273883342743, + 0.03474779799580574, + 0.03970634937286377, + 0.026828447356820107, + 0.03020956739783287, + 0.008343657478690147, + 0.0062275053933262825, + 0.01190780010074377, + 0.006352001801133156, + 0.007981142960488796, + 0.007778808008879423, + 0.006125899963080883, + 0.007804053369909525, + 0.0058658490888774395, + 0.003369983285665512, + 0.009540272876620293, + 0.005230613984167576, + 0.004229913931339979, + 0.01601424440741539, + 0.005201671272516251, + 0.01265607587993145, + 0.0033727262634783983, + 0.0020882957614958286, + 0.0034061751794070005, + 0.0024972266983240843, + 0.0028836275450885296, + 0.0030557529535144567, + 0.0025617373175919056, + 0.0032221886795014143, + 0.0013472756836563349, + 0.001616144087165594, + 0.0021168990060687065, + 0.003415673738345504, + 0.0012087838258594275, + 0.0018245254177600145, + 0.0023665199987590313, + 0.0025445264764130116, + 0.005075876135379076, + 0.0031061081681400537, + 0.003995067905634642, + 0.003054047003388405, + 0.0035750528331846, + 0.004073399119079113, + 0.004323202650994062, + 0.003655492328107357, + 0.004490855149924755, + 0.0054339030757546425, + 0.004172886721789837, + 0.006161288823932409, + 0.003983764909207821, + 0.005638505797833204, + 0.005746340379118919, + 0.005372256971895695, + 0.005735455080866814, + 0.005328043829649687, + 0.004852824378758669, + 0.007200907915830612, + 0.007986839860677719, + 0.005863781087100506, + 0.005772722885012627, + 0.007401879411190748, + 0.029493387788534164, + 0.01494050957262516, + 0.13580547273159027, + 0.05502059683203697, + 0.020284581929445267, + 0.09597351402044296, + 0.03851180151104927, + 0.020053880289196968, + 0.006683660205453634, + 0.008802891708910465, + 0.007560958154499531, + 0.01259259507060051, + 0.006709292996674776, + 0.010895933955907822, + 0.01184028945863247, + 0.006523022428154945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.011083701625466347 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864569380879402 + ], + "opposite_logprob_tokens": [ + -3.0486457347869873 + ], + "logprob_reward": -0.04864569380879402, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864569380879402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25285112857818604, + 0.04171329364180565, + 0.06155993789434433, + 0.058283936232328415, + 0.032442495226860046, + 0.040360476821660995, + 0.056184522807598114, + 0.04326128214597702, + 0.15723875164985657, + 0.25610417127609253 + ], + "credit_map": [ + 0.022561822086572647, + 0.018509015440940857, + 0.01568794623017311, + 0.013675030320882797, + 0.05230594426393509, + 0.07119007408618927, + 0.019119255244731903, + 0.039802055805921555, + 0.0047968109138309956, + 0.0043845814652740955, + 0.008102658204734325, + 0.004135891795158386, + 0.004974935669451952, + 0.004557830281555653, + 0.004949029069393873, + 0.005811557173728943, + 0.006279739085584879, + 0.002732478315010667, + 0.006533605046570301, + 0.003919858485460281, + 0.004732774570584297, + 0.01894473284482956, + 0.004322280641645193, + 0.014094466343522072, + 0.005412903614342213, + 0.004563987720757723, + 0.008334213867783546, + 0.00525481253862381, + 0.006362853571772575, + 0.012522437609732151, + 0.006386366672813892, + 0.009446362964808941, + 0.0030673122964799404, + 0.003409304888918996, + 0.004165919963270426, + 0.0036895398516207933, + 0.004175498150289059, + 0.0038255727849900723, + 0.005741572473198175, + 0.004367774818092585, + 0.004841542337089777, + 0.003430753480643034, + 0.005742859560996294, + 0.006165368482470512, + 0.003409731900319457, + 0.004854434169828892, + 0.006090302020311356, + 0.005825482774525881, + 0.006154404021799564, + 0.005434746388345957, + 0.0057090213522315025, + 0.009263642132282257, + 0.005815534852445126, + 0.007122129667550325, + 0.007101614028215408, + 0.009583424776792526, + 0.004740774165838957, + 0.0065905433148145676, + 0.00402069604024291, + 0.007504869718104601, + 0.00532235624268651, + 0.004850993398576975, + 0.004874532576650381, + 0.005356518551707268, + 0.01888447441160679, + 0.010059988126158714, + 0.03847096487879753, + 0.0154042923822999, + 0.009924503043293953, + 0.03493756055831909, + 0.015368452295660973, + 0.014188500121235847, + 0.02230108343064785, + 0.02268366701900959, + 0.047018617391586304, + 0.03618190810084343, + 0.024823784828186035, + 0.05973971635103226, + 0.02597757801413536, + 0.017377831041812897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.04864569380879402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493861436843872 + ], + "opposite_logprob_tokens": [ + -2.929938554763794 + ], + "logprob_reward": -0.05493861436843872, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493861436843872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27285030484199524, + 0.0661226212978363, + 0.12887151539325714, + 0.08130927383899689, + 0.024121787399053574, + 0.07826250791549683, + 0.07939031720161438, + 0.04645080491900444, + 0.1660654991865158, + 0.056555356830358505 + ], + "credit_map": [ + 0.026973670348525047, + 0.029175080358982086, + 0.03277244418859482, + 0.01861918531358242, + 0.03427286073565483, + 0.05813886970281601, + 0.029700271785259247, + 0.04319791868329048, + 0.007754282560199499, + 0.006719219032675028, + 0.011256130412220955, + 0.006780985277146101, + 0.0057404921390116215, + 0.011098953895270824, + 0.007263687904924154, + 0.009508876129984856, + 0.013703588396310806, + 0.005528387147933245, + 0.013422434218227863, + 0.007719798479229212, + 0.008444062434136868, + 0.04181290790438652, + 0.007852130569517612, + 0.030388223007321358, + 0.0074430122040212154, + 0.005305973347276449, + 0.014102560468018055, + 0.011064441874623299, + 0.005844748113304377, + 0.014243300072848797, + 0.009486867114901543, + 0.01381837297230959, + 0.0032668577041476965, + 0.0029625401366502047, + 0.002773123327642679, + 0.003315771697089076, + 0.002998760901391506, + 0.002549511846154928, + 0.0034319099504500628, + 0.00282331183552742, + 0.01182533148676157, + 0.008322199806571007, + 0.010424275882542133, + 0.007864013314247131, + 0.010395081713795662, + 0.01137431152164936, + 0.009657462127506733, + 0.008399838581681252, + 0.008752263151109219, + 0.00765872560441494, + 0.00982359703630209, + 0.012443372048437595, + 0.008789355866611004, + 0.009082717821002007, + 0.009486685506999493, + 0.01335360947996378, + 0.00625529745593667, + 0.004206288140267134, + 0.004804680123925209, + 0.003906645812094212, + 0.0065738363191485405, + 0.005889641586691141, + 0.005897118244320154, + 0.008917303755879402, + 0.022778360173106194, + 0.016449162736535072, + 0.03131456300616264, + 0.016523685306310654, + 0.016119228675961494, + 0.01937079057097435, + 0.023666296154260635, + 0.019843412563204765, + 0.0036732694134116173, + 0.005525815766304731, + 0.010665074922144413, + 0.008200787007808685, + 0.00465759914368391, + 0.014255638234317303, + 0.006496133748441935, + 0.0030810420867055655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.05493861436843872 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a steady tracking shot from a slightly upward angle, capturing a white egret soaring over a vast expanse of lush reeds swaying gently in the breeze. The egret's slender figure, its flawless snow-white feathers appearing remarkably pure in the soft, diffused morning light. Its broad wings flap gracefully, its slender black legs tucked back, its sharp yellow beak pointed forward, its movements fluid and deliberate. The world is enveloped in a hazy, milky morning mist. A veil-like haze rises from the tranquil surface of the lake, blurring distant scenery and creating an ethereal, tranquil atmosphere. The camera then steadily and smoothly pulls back, gradually reducing the egret's size to a nimble white dot. Its flight path traces a graceful arc across the vast canvas of golden reeds and pale gray sky, revealing a serene and majestic panorama of the wetland.", + "seed": 18377, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_13/index_365_seed_18377_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03190960604697466, + "vlm_logprob": -0.03190960604697466, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03190960604697466, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011463016271591187 + ], + "opposite_logprob_tokens": [ + -4.761463165283203 + ], + "logprob_reward": -0.011463016271591187, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011463016271591187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4441331923007965, + 0.10787256062030792, + 0.03567592427134514, + 0.05403772369027138, + 0.03491303324699402, + 0.058439385145902634, + 0.132070854306221, + 0.08652988076210022, + 0.031143944710493088, + 0.015183502808213234 + ], + "credit_map": [ + 0.053698644042015076, + 0.05426975339651108, + 0.06472522020339966, + 0.051380034536123276, + 0.0495292954146862, + 0.051857780665159225, + 0.04862337186932564, + 0.07004906982183456, + 0.008711203932762146, + 0.013468848541378975, + 0.013725955039262772, + 0.013971241191029549, + 0.00990300439298153, + 0.01466158777475357, + 0.013494627550244331, + 0.0199360903352499, + 0.0030220584012567997, + 0.003978724591434002, + 0.0043563563376665115, + 0.004690989386290312, + 0.0050684018060564995, + 0.005203752778470516, + 0.004942420870065689, + 0.004413219168782234, + 0.005296709481626749, + 0.007464906200766563, + 0.0050427294336259365, + 0.008094503544270992, + 0.007937818765640259, + 0.007752167992293835, + 0.008621977642178535, + 0.0038269045762717724, + 0.0031880096066743135, + 0.004218905698508024, + 0.005409695208072662, + 0.004644396714866161, + 0.0038220160640776157, + 0.004340414423495531, + 0.004842970985919237, + 0.004446623846888542, + 0.004289559554308653, + 0.007481250446289778, + 0.006086322478950024, + 0.0040381355211138725, + 0.00819095503538847, + 0.0071776616387069225, + 0.011525413952767849, + 0.00965008232742548, + 0.0178044606000185, + 0.019396966323256493, + 0.00840330496430397, + 0.010223142802715302, + 0.0336645245552063, + 0.017431460320949554, + 0.011649510823190212, + 0.013497479259967804, + 0.00493613863363862, + 0.029019687324762344, + 0.003738246625289321, + 0.004425931256264448, + 0.02865798957645893, + 0.0053212703205645084, + 0.004285700153559446, + 0.006144917104393244, + 0.0037498394958674908, + 0.00660740165039897, + 0.0018079489236697555, + 0.0019367614295333624, + 0.0064876023679971695, + 0.0059183817356824875, + 0.0020799729973077774, + 0.002556032966822386, + 0.0029453050810843706, + 0.0022154536563903093, + 0.0013323001330718398, + 0.0012082557659596205, + 0.0018823628779500723, + 0.002691612346097827, + 0.0013389274245128036, + 0.0015692844754084945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.011463016271591187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635142020881176 + ], + "opposite_logprob_tokens": [ + -3.6513514518737793 + ], + "logprob_reward": -0.02635142020881176, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635142020881176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3675191402435303, + 0.11036340892314911, + 0.043583884835243225, + 0.016654400154948235, + 0.08971459418535233, + 0.2194065898656845, + 0.0629512146115303, + 0.031472090631723404, + 0.03925338760018349, + 0.01908128522336483 + ], + "credit_map": [ + 0.044909317046403885, + 0.04510033130645752, + 0.05691960081458092, + 0.041482072323560715, + 0.04039754718542099, + 0.04292020946741104, + 0.03655945137143135, + 0.05923062190413475, + 0.009528817608952522, + 0.01238312292844057, + 0.016678521409630775, + 0.016517551615834236, + 0.00896085612475872, + 0.012489616870880127, + 0.017557509243488312, + 0.016247417777776718, + 0.0037841047160327435, + 0.006364443339407444, + 0.0036010313779115677, + 0.004378355573862791, + 0.009169499389827251, + 0.007328372914344072, + 0.005137478932738304, + 0.0038205995224416256, + 0.0018247633706778288, + 0.0022646256256848574, + 0.002293061465024948, + 0.00314431288279593, + 0.0012701281812041998, + 0.0013919260818511248, + 0.002032305346801877, + 0.002433277200907469, + 0.009828243404626846, + 0.011573759838938713, + 0.01242730300873518, + 0.013520631939172745, + 0.008080664090812206, + 0.011554951779544353, + 0.013116948306560516, + 0.00961209274828434, + 0.01939350925385952, + 0.0259636789560318, + 0.039500560611486435, + 0.021428685635328293, + 0.016820577904582024, + 0.028759198263287544, + 0.02569803223013878, + 0.041842348873615265, + 0.007052262779325247, + 0.00824342854321003, + 0.005748233292251825, + 0.009290266782045364, + 0.012207688763737679, + 0.008492524735629559, + 0.006448648404330015, + 0.005468160379678011, + 0.003991943784058094, + 0.0047105057165026665, + 0.004710973706096411, + 0.004710379522293806, + 0.004284584429115057, + 0.002827680902555585, + 0.003505064407363534, + 0.0027309590950608253, + 0.005815483629703522, + 0.009952006861567497, + 0.0030756783671677113, + 0.003123497124761343, + 0.00960581935942173, + 0.0027242102660238743, + 0.0022080258931964636, + 0.002748664701357484, + 0.0018232794245705009, + 0.0027478470001369715, + 0.0024984569754451513, + 0.0031481850892305374, + 0.0019530582940205932, + 0.0029859659262001514, + 0.001950673758983612, + 0.0019738185219466686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02635142020881176 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079574935138226 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079574935138226, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079574935138226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.791203498840332, + 0.07406803220510483, + 0.029507577419281006, + 0.008649460040032864, + 0.005874255206435919, + 0.01610609143972397, + 0.03587600961327553, + 0.01671025902032852, + 0.014729849994182587, + 0.007274923380464315 + ], + "credit_map": [ + 0.09172601997852325, + 0.0984310656785965, + 0.11692183464765549, + 0.0879518985748291, + 0.09277886897325516, + 0.09813961386680603, + 0.0782882571220398, + 0.12696599960327148, + 0.006276994943618774, + 0.008161570876836777, + 0.012379249557852745, + 0.01255745068192482, + 0.006800501141697168, + 0.00802568905055523, + 0.012506477534770966, + 0.0073601012118160725, + 0.002478096866980195, + 0.004204191733151674, + 0.0027520558796823025, + 0.0033443295396864414, + 0.006778296548873186, + 0.0035859085619449615, + 0.003546975553035736, + 0.0028177218046039343, + 0.0011689391685649753, + 0.0012233455199748278, + 0.0007721426081843674, + 0.001261040335521102, + 0.0013170482125133276, + 0.0010690444614738226, + 0.0010250936029478908, + 0.0008128058398142457, + 0.000701284094247967, + 0.0006743810954503715, + 0.0006768150487914681, + 0.00073794781928882, + 0.0005421106470748782, + 0.0007780538289807737, + 0.0009524181368760765, + 0.0008112444775179029, + 0.001538631389848888, + 0.0021474575623869896, + 0.0027066476177424192, + 0.0014270310057327151, + 0.002249774057418108, + 0.0018945287447422743, + 0.0029169267509132624, + 0.001225094310939312, + 0.004821299109607935, + 0.00536347646266222, + 0.002211431274190545, + 0.002492582192644477, + 0.009371073916554451, + 0.0051277149468660355, + 0.0030113679822534323, + 0.0034770627971738577, + 0.0012497910065576434, + 0.004370512440800667, + 0.0011845643166452646, + 0.0010771879460662603, + 0.0042725359089672565, + 0.001764538697898388, + 0.0010503024095669389, + 0.001740825711749494, + 0.001121027860790491, + 0.003959171939641237, + 0.0006446664920076728, + 0.0009595228475518525, + 0.004198824055492878, + 0.0019315119134262204, + 0.0009495108970440924, + 0.0009656139300204813, + 0.001033071312122047, + 0.001312830368988216, + 0.0005306603270582855, + 0.0005413081380538642, + 0.0012724846601486206, + 0.0015555296558886766, + 0.00047074799658730626, + 0.0005582909216172993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.011079574935138226 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048631615936756134 + ], + "opposite_logprob_tokens": [ + -3.0486316680908203 + ], + "logprob_reward": -0.048631615936756134, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048631615936756134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41792556643486023, + 0.1109364777803421, + 0.04663568362593651, + 0.1555042266845703, + 0.028988225385546684, + 0.06564097106456757, + 0.06468673050403595, + 0.05485498905181885, + 0.029861362650990486, + 0.02496568113565445 + ], + "credit_map": [ + 0.043723490089178085, + 0.049737703055143356, + 0.0626673698425293, + 0.05517300218343735, + 0.04392208158969879, + 0.048323750495910645, + 0.048512738198041916, + 0.06586548686027527, + 0.009562227874994278, + 0.013442217372357845, + 0.014281426556408405, + 0.012353513389825821, + 0.00939800962805748, + 0.017628077417612076, + 0.016652261838316917, + 0.017618747428059578, + 0.004312468692660332, + 0.006395863369107246, + 0.004021940287202597, + 0.00486403563991189, + 0.009126603603363037, + 0.006829930003732443, + 0.005369964521378279, + 0.00571488169953227, + 0.019801199436187744, + 0.026376329362392426, + 0.014982865191996098, + 0.019500836730003357, + 0.009220768697559834, + 0.016229772940278053, + 0.020414402708411217, + 0.02897806465625763, + 0.0025332754012197256, + 0.003485436085611582, + 0.005843728315085173, + 0.004922190681099892, + 0.002000868786126375, + 0.0021717955823987722, + 0.004687017761170864, + 0.0033439146354794502, + 0.005799990147352219, + 0.007714611012488604, + 0.014261765405535698, + 0.006505392491817474, + 0.00501689687371254, + 0.007587181869894266, + 0.00732424296438694, + 0.011430894024670124, + 0.008983170613646507, + 0.010890156961977482, + 0.004264055751264095, + 0.006474028341472149, + 0.015381982550024986, + 0.006345342379063368, + 0.0059930006973445415, + 0.006355004385113716, + 0.0055493987165391445, + 0.012740994803607464, + 0.004500985611230135, + 0.005843106657266617, + 0.01259637251496315, + 0.004102480132132769, + 0.00452582398429513, + 0.004995835479348898, + 0.0025467034429311752, + 0.007067760918289423, + 0.0014770468696951866, + 0.0019646298605948687, + 0.007065993268042803, + 0.005729213356971741, + 0.0019465715158730745, + 0.002063445281237364, + 0.0032493427861481905, + 0.004440345801413059, + 0.0017826722469180822, + 0.0017027414869517088, + 0.003313853172585368, + 0.00627045426517725, + 0.0018560434691607952, + 0.002350230235606432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.048631615936756134 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202240288257599 + ], + "opposite_logprob_tokens": [ + -2.8120224475860596 + ], + "logprob_reward": -0.06202240288257599, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202240288257599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08299532532691956, + 0.09805571287870407, + 0.04805338382720947, + 0.07973863184452057, + 0.07806876301765442, + 0.05770539864897728, + 0.22030958533287048, + 0.27509161829948425, + 0.0346616692841053, + 0.02531987801194191 + ], + "credit_map": [ + 0.0065622185356915, + 0.010282247327268124, + 0.009625716134905815, + 0.01309999730437994, + 0.012237105518579483, + 0.008925572037696838, + 0.01182635873556137, + 0.010436112061142921, + 0.007250220514833927, + 0.011934008449316025, + 0.009484407491981983, + 0.011080746538937092, + 0.008745208382606506, + 0.017517898231744766, + 0.017176277935504913, + 0.014866945333778858, + 0.005004270933568478, + 0.007632289547473192, + 0.0045911590568721294, + 0.004124082159250975, + 0.011626075021922588, + 0.005730975419282913, + 0.0048806252889335155, + 0.004463904537260532, + 0.008642804808914661, + 0.011056485585868359, + 0.00918414443731308, + 0.013244039379060268, + 0.005925408564507961, + 0.006701970007270575, + 0.009060351178050041, + 0.015923431143164635, + 0.007697924505919218, + 0.009782806038856506, + 0.007598450873047113, + 0.006258170586079359, + 0.010540558956563473, + 0.013874092139303684, + 0.010017517022788525, + 0.012299244292080402, + 0.0053697084076702595, + 0.008042418397963047, + 0.00928987842053175, + 0.004196838941425085, + 0.008412608876824379, + 0.01019534096121788, + 0.006926974281668663, + 0.005271628964692354, + 0.028613323345780373, + 0.03142562136054039, + 0.014292691834270954, + 0.017068374902009964, + 0.05494041368365288, + 0.03125857934355736, + 0.019998498260974884, + 0.022712090983986855, + 0.013563890941441059, + 0.09671686589717865, + 0.00923741701990366, + 0.00982091948390007, + 0.09506344050168991, + 0.018413031473755836, + 0.011772245168685913, + 0.020503820851445198, + 0.0037161810323596, + 0.006048085168004036, + 0.0018829243490472436, + 0.002243304392322898, + 0.006124462932348251, + 0.008617176674306393, + 0.0019057326717302203, + 0.004123800899833441, + 0.0023349702823907137, + 0.005593536421656609, + 0.0015861610881984234, + 0.0012850796338170767, + 0.004933908116072416, + 0.005492411553859711, + 0.0017616533441469073, + 0.002332156989723444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.06202240288257599 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a steady tracking shot from a slightly upward angle, capturing a white egret soaring over a vast expanse of lush reeds swaying gently in the breeze. The egret's slender figure, its flawless snow-white feathers appearing remarkably pure in the soft, diffused morning light. Its broad wings flap gracefully, its slender black legs tucked back, its sharp yellow beak pointed forward, its movements fluid and deliberate. The world is enveloped in a hazy, milky morning mist. A veil-like haze rises from the tranquil surface of the lake, blurring distant scenery and creating an ethereal, tranquil atmosphere. The camera then steadily and smoothly pulls back, gradually reducing the egret's size to a nimble white dot. Its flight path traces a graceful arc across the vast canvas of golden reeds and pale gray sky, revealing a serene and majestic panorama of the wetland.", + "seed": 18378, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_13/index_365_seed_18378_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05684674512594938, + "vlm_logprob": -0.05684674512594938, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.05684674512594938, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015059704892337322 + ], + "opposite_logprob_tokens": [ + -4.640059471130371 + ], + "logprob_reward": -0.015059704892337322, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.015059704892337322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2268797755241394, + 0.08070207387208939, + 0.07697174698114395, + 0.15496328473091125, + 0.10592975467443466, + 0.08114909380674362, + 0.12942427396774292, + 0.04912634938955307, + 0.04862498119473457, + 0.046228621155023575 + ], + "credit_map": [ + 0.0183828454464674, + 0.00842099729925394, + 0.08363249152898788, + 0.009363255463540554, + 0.012532869353890419, + 0.015983566641807556, + 0.02800382487475872, + 0.05055991932749748, + 0.006847454700618982, + 0.00813302118331194, + 0.013014202937483788, + 0.010219855234026909, + 0.0071366531774401665, + 0.00988719705492258, + 0.01338509563356638, + 0.012078595347702503, + 0.008631475269794464, + 0.007793395780026913, + 0.009992429986596107, + 0.013079791329801083, + 0.007415327243506908, + 0.012178730219602585, + 0.010625130496919155, + 0.00725546246394515, + 0.011894849129021168, + 0.007828781381249428, + 0.05043995380401611, + 0.022763563320040703, + 0.008762545883655548, + 0.018950168043375015, + 0.014172676019370556, + 0.02015073411166668, + 0.012345577590167522, + 0.01462776493281126, + 0.013524441979825497, + 0.015616502612829208, + 0.013352366164326668, + 0.015966372564435005, + 0.011175712570548058, + 0.009321009740233421, + 0.008072170428931713, + 0.01211407408118248, + 0.009178750216960907, + 0.01187343243509531, + 0.008551225997507572, + 0.01098515372723341, + 0.012577293440699577, + 0.007796990219503641, + 0.007699375972151756, + 0.005672556813806295, + 0.04188762232661247, + 0.006361974868923426, + 0.0066955676302313805, + 0.009726115502417088, + 0.0423605814576149, + 0.009020467288792133, + 0.005446822382509708, + 0.00460898969322443, + 0.007900563068687916, + 0.007741330657154322, + 0.005064796190708876, + 0.00549623928964138, + 0.008545517921447754, + 0.00432208925485611, + 0.005513547919690609, + 0.005948739126324654, + 0.006064400542527437, + 0.005342013668268919, + 0.00514708086848259, + 0.009199599735438824, + 0.0051546297036111355, + 0.006254970096051693, + 0.0065820487216115, + 0.0048917862586677074, + 0.004882815293967724, + 0.0056078629568219185, + 0.005492737982422113, + 0.006642826367169619, + 0.005299303214997053, + 0.0068292440846562386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.015059704892337322 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981729805469513 + ], + "opposite_logprob_tokens": [ + -3.5298173427581787 + ], + "logprob_reward": -0.02981729805469513, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981729805469513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7923452258110046, + 0.026599403470754623, + 0.0101223886013031, + 0.07176472246646881, + 0.01095344964414835, + 0.009026440791785717, + 0.03270774707198143, + 0.018979761749505997, + 0.00939315464347601, + 0.018107615411281586 + ], + "credit_map": [ + 0.03780591860413551, + 0.008785752579569817, + 0.3397439420223236, + 0.017943179234862328, + 0.017809586599469185, + 0.03343581780791283, + 0.054955244064331055, + 0.2818658947944641, + 0.002741915872320533, + 0.0026193216908723116, + 0.0036015089135617018, + 0.002612298121675849, + 0.0026009585708379745, + 0.004904544446617365, + 0.004258157219737768, + 0.0032607046887278557, + 0.0011968952603638172, + 0.0010038686450570822, + 0.0013936379691585898, + 0.0011504960712045431, + 0.0009824538137763739, + 0.0014809111598879099, + 0.0015194826992228627, + 0.0013946443796157837, + 0.0026629928033798933, + 0.0013974568573758006, + 0.023865319788455963, + 0.0054579549469053745, + 0.0020811434369534254, + 0.003152362070977688, + 0.028906503692269325, + 0.004240993410348892, + 0.0008984830928966403, + 0.0011950453044846654, + 0.0015012224903330207, + 0.0015310738235712051, + 0.0013536937767639756, + 0.0012751322938129306, + 0.0015776752261444926, + 0.0016211244510486722, + 0.0008443415863439441, + 0.0008651940734125674, + 0.001210098504088819, + 0.0013651930494233966, + 0.0008462590631097555, + 0.0014555825619027019, + 0.0010855169966816902, + 0.0013542560627683997, + 0.0015288111753761768, + 0.0010497898329049349, + 0.011963769793510437, + 0.0014031665632501245, + 0.0013950420543551445, + 0.0017384556122124195, + 0.01154891774058342, + 0.0020797965116798878, + 0.0021456973627209663, + 0.002066150074824691, + 0.0032046244014054537, + 0.0026799484621733427, + 0.002130849752575159, + 0.002541206544265151, + 0.002656947821378708, + 0.0015543404733762145, + 0.001048654899932444, + 0.0009421640424989164, + 0.001208737026900053, + 0.0010619561653584242, + 0.0008449330343864858, + 0.0017483447445556521, + 0.0010980080114677548, + 0.0014403578825294971, + 0.0029562502168118954, + 0.0011876840144395828, + 0.0032514629419893026, + 0.0015843595610931516, + 0.0016247146995738149, + 0.002263539470732212, + 0.0019078991608694196, + 0.0033317068591713905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.02981729805469513 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308921605348587 + ], + "opposite_logprob_tokens": [ + -3.7733089923858643 + ], + "logprob_reward": -0.023308921605348587, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308921605348587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5264189839363098, + 0.026241270825266838, + 0.014444388449192047, + 0.29027482867240906, + 0.03178098425269127, + 0.02988525666296482, + 0.022738449275493622, + 0.010549947619438171, + 0.012192953377962112, + 0.035472940653562546 + ], + "credit_map": [ + 0.025322116911411285, + 0.007128700613975525, + 0.22361209988594055, + 0.013151472434401512, + 0.011714036576449871, + 0.022327253594994545, + 0.0357045941054821, + 0.18745869398117065, + 0.0015948716318234801, + 0.001806313288398087, + 0.005019147414714098, + 0.0030570554081350565, + 0.001985774142667651, + 0.0028049449902027845, + 0.005848839413374662, + 0.0041243233717978, + 0.0016472423449158669, + 0.001850440981797874, + 0.0016136596677824855, + 0.0018426033202558756, + 0.0015065716579556465, + 0.0021176591981202364, + 0.0014520069817081094, + 0.0024142039474099874, + 0.012139503844082355, + 0.003175899386405945, + 0.1176113560795784, + 0.014200829900801182, + 0.006328646093606949, + 0.012897108681499958, + 0.11104404181241989, + 0.012877456843852997, + 0.002887881360948086, + 0.003801287617534399, + 0.00432828301563859, + 0.005119671113789082, + 0.003950163256376982, + 0.0039729890413582325, + 0.004606772214174271, + 0.0031139382626861334, + 0.0036920991260558367, + 0.00369048910215497, + 0.0033046910539269447, + 0.004440088756382465, + 0.0027005095034837723, + 0.005174342542886734, + 0.003944390453398228, + 0.0029386465903371572, + 0.001077177468687296, + 0.0008660599705763161, + 0.0077289738692343235, + 0.0012938561849296093, + 0.0012031474616378546, + 0.001497495686635375, + 0.007555922958999872, + 0.0015158153837546706, + 0.0011500135296955705, + 0.0011356734903529286, + 0.001359294750727713, + 0.0014375124592334032, + 0.0010317668784409761, + 0.000971259898506105, + 0.0017869756557047367, + 0.00167745107319206, + 0.0012592088896781206, + 0.0012145870132371783, + 0.0014859455404803157, + 0.0015062636230140924, + 0.00139811378903687, + 0.0028422395698726177, + 0.0015417614486068487, + 0.0009448325727134943, + 0.005204902030527592, + 0.003041315358132124, + 0.00577158248052001, + 0.0034640806261450052, + 0.0033986885100603104, + 0.005006845574826002, + 0.004317205399274826, + 0.005268322769552469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.023308921605348587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903313428163528 + ], + "opposite_logprob_tokens": [ + -2.4640331268310547 + ], + "logprob_reward": -0.08903313428163528, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903313428163528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7342895865440369, + 0.04458821937441826, + 0.014742549508810043, + 0.0690116435289383, + 0.02428666315972805, + 0.006004721857607365, + 0.0222718995064497, + 0.04447271674871445, + 0.021245364099740982, + 0.019086631014943123 + ], + "credit_map": [ + 0.03530533239245415, + 0.008090421557426453, + 0.31475386023521423, + 0.0167255662381649, + 0.0163862444460392, + 0.03156544268131256, + 0.0524253211915493, + 0.2590373754501343, + 0.004013278521597385, + 0.002627381356433034, + 0.007407406345009804, + 0.004642937798053026, + 0.004244191572070122, + 0.0057647209614515305, + 0.009578447788953781, + 0.006309847347438335, + 0.0013482437934726477, + 0.0014023793628439307, + 0.002039611106738448, + 0.0021083103492856026, + 0.0013555645709857345, + 0.002134207170456648, + 0.0027167596854269505, + 0.001637471723370254, + 0.0030698291957378387, + 0.002289213938638568, + 0.027554847300052643, + 0.0051813856698572636, + 0.0023892270401120186, + 0.004739122930914164, + 0.019824549555778503, + 0.0039634667336940765, + 0.002118875039741397, + 0.0031802523881196976, + 0.0038119156379252672, + 0.0037455051206052303, + 0.0021876811515539885, + 0.0030642347410321236, + 0.003922189120203257, + 0.0022560080979019403, + 0.0005568669876083732, + 0.0004595788777805865, + 0.0007488176343031228, + 0.0005240049795247614, + 0.000591421325225383, + 0.001099183689802885, + 0.0010726155014708638, + 0.0009522326290607452, + 0.0012096228310838342, + 0.0008100876002572477, + 0.007281000260263681, + 0.0010549642611294985, + 0.0011002783430740237, + 0.001595691661350429, + 0.00739625608548522, + 0.0018239960772916675, + 0.0031508069951087236, + 0.00477731553837657, + 0.0054940637201070786, + 0.007047367747873068, + 0.0037447609938681126, + 0.005978346802294254, + 0.00809065904468298, + 0.006189391016960144, + 0.0018915648106485605, + 0.0014230197994038463, + 0.0024008657783269882, + 0.003814132185652852, + 0.0014402872184291482, + 0.002724407473579049, + 0.003975282423198223, + 0.0035758016165345907, + 0.0033627119846642017, + 0.001714550075121224, + 0.0024893665686249733, + 0.0017770567210391164, + 0.0017761854687705636, + 0.0021902609150856733, + 0.0019937846809625626, + 0.0037827114574611187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.08903313428163528 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a white egret with snow-white feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the egret soaring over a vast expanse of lush reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the egret have a sharp yellow beak and slender black legs tucked back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment characterized by a hazy, milky morning mist over a lake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a steady tracking shot from a slightly upward angle and then pull back smoothly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270146667957306 + ], + "opposite_logprob_tokens": [ + -2.127014636993408 + ], + "logprob_reward": -0.1270146667957306, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270146667957306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37993916869163513, + 0.03679264709353447, + 0.01709173433482647, + 0.3056497871875763, + 0.03659971058368683, + 0.020929906517267227, + 0.11026318371295929, + 0.028137095272541046, + 0.01507593598216772, + 0.049520883709192276 + ], + "credit_map": [ + 0.01903744973242283, + 0.011641476303339005, + 0.15405787527561188, + 0.013404223136603832, + 0.012043150141835213, + 0.019924310967326164, + 0.029544251039624214, + 0.1202864795923233, + 0.0025285398587584496, + 0.0040402403101325035, + 0.0053382087498903275, + 0.00549312774091959, + 0.0031798037234693766, + 0.00510089797899127, + 0.006323291454464197, + 0.004788537509739399, + 0.00150718679651618, + 0.0021664053201675415, + 0.0024258121848106384, + 0.0030354969203472137, + 0.0018245724495500326, + 0.001956216525286436, + 0.002319600200280547, + 0.0018564441706985235, + 0.012481301091611385, + 0.005108910612761974, + 0.10885410010814667, + 0.01816527545452118, + 0.008334875106811523, + 0.014571350067853928, + 0.12241166830062866, + 0.01572231575846672, + 0.002786524360999465, + 0.003146337578073144, + 0.005168092902749777, + 0.007733548060059547, + 0.004663463216274977, + 0.004522660281509161, + 0.004186846781522036, + 0.004392236936837435, + 0.002318291924893856, + 0.0024249516427516937, + 0.0023812083527445793, + 0.002740565687417984, + 0.0018772536423057318, + 0.002895364537835121, + 0.003872356377542019, + 0.0024199134204536676, + 0.005749458447098732, + 0.0046700467355549335, + 0.03778763487935066, + 0.004817407578229904, + 0.005664324853569269, + 0.006911748554557562, + 0.03699784353375435, + 0.0076647186651825905, + 0.0021699306089431047, + 0.002930891001597047, + 0.004097694996744394, + 0.005366665776818991, + 0.0028759005945175886, + 0.00206015701405704, + 0.005900864489376545, + 0.002734990557655692, + 0.001458877231925726, + 0.002110905945301056, + 0.0018302202224731445, + 0.0016020763432607055, + 0.001816746429540217, + 0.002331797732040286, + 0.0023599951528012753, + 0.0015653163427487016, + 0.005107589066028595, + 0.006501640193164349, + 0.006842796225100756, + 0.007077592890709639, + 0.006333388388156891, + 0.005099655129015446, + 0.005840359255671501, + 0.006717863492667675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.1270146667957306 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_14/index_tensor([7508, 7508])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_14/index_tensor([7508, 7508])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4354f6c6dee8d2ab7f68ff3994de1688c27d10b1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_14/index_tensor([7508, 7508])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, first-person close-up captures a clean hand. The tips of its thumb and index finger delicately grasp a few shrunken, yet vibrant red goji berries and several whole, dried chrysanthemums with pale yellow petals, each as delicate as paper. Soft afternoon sunlight streams through the side window, casting a warm golden glow over the cozy kitchen and illuminating the tiny dust particles floating in the air. The hand steadily and deliberately moves directly above a crystal-clear, round glass teapot. The sunlight reflects off the smooth, curved surface, casting dazzling, warm spots of light. The warm wooden countertop and a blurred kitchen corner in the background further enhance the homely atmosphere. As the fingertips gently release their grip, the red goji berries and pale yellow chrysanthemums drift gracefully in slow motion, finally scattering silently on the polished bottom of the pot. The entire action imbued with the tranquil ritual of preparing a cup of warming herbal tea, creating an atmosphere of ultimate relaxation, healing, and a sense of tranquility.", + "seed": 19001, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_14/index_7508_seed_19001_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021894399635493757, + "vlm_logprob": -0.021894399635493757, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.021894399635493757, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028380686417222023 + ], + "opposite_logprob_tokens": [ + -4.278380870819092 + ], + "logprob_reward": -0.028380686417222023, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.028380686417222023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12956254184246063, + 0.059441909193992615, + 0.10632214695215225, + 0.21545347571372986, + 0.03488091379404068, + 0.04337478056550026, + 0.05117952451109886, + 0.14980602264404297, + 0.15979988873004913, + 0.050178807228803635 + ], + "credit_map": [ + 0.021678509190678596, + 0.009015333838760853, + 0.02195916511118412, + 0.008512916974723339, + 0.008959420956671238, + 0.037709034979343414, + 0.009336070157587528, + 0.012392084114253521, + 0.004154518712311983, + 0.0039877137169241905, + 0.006607730872929096, + 0.007517991587519646, + 0.006548399105668068, + 0.008130772039294243, + 0.008617249317467213, + 0.013877532444894314, + 0.010087361559271812, + 0.009804264642298222, + 0.012384984642267227, + 0.005564116407185793, + 0.007907683961093426, + 0.02203148603439331, + 0.007079351227730513, + 0.0314629003405571, + 0.021979227662086487, + 0.01619778759777546, + 0.0277346670627594, + 0.023708945140242577, + 0.017942853271961212, + 0.02450837381184101, + 0.0283355750143528, + 0.055046048015356064, + 0.00269131432287395, + 0.0035331116523593664, + 0.0032056868076324463, + 0.0031671517062932253, + 0.00903205294162035, + 0.004184078890830278, + 0.0033715167082846165, + 0.005695997271686792, + 0.0035951926838606596, + 0.005860781297087669, + 0.004033098462969065, + 0.0027080916333943605, + 0.00878221821039915, + 0.0059717921540141106, + 0.0051203020848333836, + 0.007303305435925722, + 0.005465642549097538, + 0.01227143406867981, + 0.0053690457716584206, + 0.0035939894150942564, + 0.00892573781311512, + 0.005959498230367899, + 0.004005997441709042, + 0.0055881766602396965, + 0.023146526888012886, + 0.01750272884964943, + 0.019087839871644974, + 0.01816180907189846, + 0.01750686578452587, + 0.012907992117106915, + 0.021303556859493256, + 0.020188694819808006, + 0.021983882412314415, + 0.019994603469967842, + 0.02222777158021927, + 0.015080772340297699, + 0.03247375786304474, + 0.017606161534786224, + 0.014547822996973991, + 0.01588512398302555, + 0.004896513652056456, + 0.0066712405532598495, + 0.006151597015559673, + 0.007062836550176144, + 0.009050488471984863, + 0.005564600229263306, + 0.004989464767277241, + 0.005792064126580954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.028380686417222023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036346344277262688 + ], + "opposite_logprob_tokens": [ + -5.628634452819824 + ], + "logprob_reward": -0.0036346344277262688, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036346344277262688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39914703369140625, + 0.14447681605815887, + 0.08654706925153732, + 0.04774902015924454, + 0.05291614681482315, + 0.04175375774502754, + 0.07015102356672287, + 0.07013890147209167, + 0.05379126965999603, + 0.03332889452576637 + ], + "credit_map": [ + 0.054213520139455795, + 0.025584889575839043, + 0.06151927635073662, + 0.026333408430218697, + 0.03958466649055481, + 0.09613342583179474, + 0.03880694881081581, + 0.056970883160829544, + 0.015171217732131481, + 0.009448491036891937, + 0.01866837590932846, + 0.014701231382787228, + 0.010289834812283516, + 0.022459756582975388, + 0.01826365850865841, + 0.0354742594063282, + 0.009904617443680763, + 0.006751172244548798, + 0.010376500897109509, + 0.0061381664127111435, + 0.005970407743006945, + 0.014977915212512016, + 0.008561307564377785, + 0.023866984993219376, + 0.003236565040424466, + 0.007974548265337944, + 0.006465418729931116, + 0.004206634126603603, + 0.007774275727570057, + 0.005981954745948315, + 0.006444161292165518, + 0.0056654601357877254, + 0.004140755161643028, + 0.009700564667582512, + 0.006916818208992481, + 0.0030955311376601458, + 0.010168945416808128, + 0.006785693112760782, + 0.005695913918316364, + 0.006411925423890352, + 0.005042476579546928, + 0.00617971969768405, + 0.0049369377084076405, + 0.0021159318275749683, + 0.010069428011775017, + 0.005666990298777819, + 0.003604843048378825, + 0.004137430340051651, + 0.007904117926955223, + 0.011210807599127293, + 0.008556636981666088, + 0.005503710359334946, + 0.010218136943876743, + 0.008010118268430233, + 0.010400451719760895, + 0.00834704376757145, + 0.01287695299834013, + 0.009380204603075981, + 0.007568646688014269, + 0.007810656912624836, + 0.007928052917122841, + 0.006363279186189175, + 0.009114379063248634, + 0.009096726775169373, + 0.004583340138196945, + 0.005278467666357756, + 0.007085742428898811, + 0.006535097490996122, + 0.009825780987739563, + 0.00612850533798337, + 0.006160707212984562, + 0.00819362886250019, + 0.00269227079115808, + 0.0038360273465514183, + 0.005001909099519253, + 0.003666339907795191, + 0.005946147255599499, + 0.004238090943545103, + 0.004270210396498442, + 0.003677898785099387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0036346344277262688 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328586019575596 + ], + "opposite_logprob_tokens": [ + -3.7732858657836914 + ], + "logprob_reward": -0.02328586019575596, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328586019575596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08953256160020828, + 0.07125595211982727, + 0.1126459464430809, + 0.20331557095050812, + 0.04641011357307434, + 0.03601428121328354, + 0.05079343914985657, + 0.2755492925643921, + 0.08455050736665726, + 0.029932307079434395 + ], + "credit_map": [ + 0.010961388237774372, + 0.006929989438503981, + 0.01583940163254738, + 0.005548939108848572, + 0.0065075973980128765, + 0.022490905597805977, + 0.00974367931485176, + 0.01151066180318594, + 0.00732868816703558, + 0.005215834826231003, + 0.01093657873570919, + 0.006090857088565826, + 0.0051386090926826, + 0.011432195082306862, + 0.009438859298825264, + 0.015674330294132233, + 0.011594261042773724, + 0.006937573663890362, + 0.013534670695662498, + 0.004796502646058798, + 0.006735887844115496, + 0.02353181131184101, + 0.008759795688092709, + 0.03675544261932373, + 0.01664465107023716, + 0.028272371739149094, + 0.02533300593495369, + 0.012847640551626682, + 0.028911031782627106, + 0.022662322968244553, + 0.028616376221179962, + 0.0400281623005867, + 0.0033205077052116394, + 0.0037965429946780205, + 0.005424462724477053, + 0.00552755780518055, + 0.007436251267790794, + 0.004987038671970367, + 0.007506720721721649, + 0.008411029353737831, + 0.003412543796002865, + 0.004771583713591099, + 0.004186300095170736, + 0.002284665359184146, + 0.006227714475244284, + 0.004340725485235453, + 0.0038299558218568563, + 0.006960793863981962, + 0.004885295405983925, + 0.008032577112317085, + 0.004745010286569595, + 0.005758184008300304, + 0.006220207549631596, + 0.005034295376390219, + 0.008677716366946697, + 0.0074401553720235825, + 0.039026178419589996, + 0.03187359869480133, + 0.026916954666376114, + 0.04643506929278374, + 0.02521844208240509, + 0.023348016664385796, + 0.0384693369269371, + 0.04426170140504837, + 0.010398119688034058, + 0.011461255140602589, + 0.014518986456096172, + 0.007717197760939598, + 0.013381414115428925, + 0.010073509067296982, + 0.007672599516808987, + 0.009327427484095097, + 0.0036826739087700844, + 0.0035834074951708317, + 0.0038669309578835964, + 0.0041748457588255405, + 0.004083361476659775, + 0.0032703073229640722, + 0.003081473521888256, + 0.004189306404441595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.02328586019575596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067780554294586 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067780554294586, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067780554294586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10396222770214081, + 0.09189023822546005, + 0.12657922506332397, + 0.13904963433742523, + 0.06610824167728424, + 0.03466149792075157, + 0.03439382463693619, + 0.29523932933807373, + 0.05887307971715927, + 0.049242664128541946 + ], + "credit_map": [ + 0.011602549813687801, + 0.011835793033242226, + 0.015065807849168777, + 0.007343323435634375, + 0.020344290882349014, + 0.012329932302236557, + 0.01526752207428217, + 0.010173022747039795, + 0.007651271298527718, + 0.005941932555288076, + 0.01150575466454029, + 0.010027192533016205, + 0.007876189425587654, + 0.013274984434247017, + 0.016732579097151756, + 0.018880341202020645, + 0.01798691786825657, + 0.009741280227899551, + 0.018435196951031685, + 0.007172806654125452, + 0.007296610623598099, + 0.019991030916571617, + 0.010648075491189957, + 0.035307325422763824, + 0.01906817965209484, + 0.019684014841914177, + 0.019622595980763435, + 0.00897306390106678, + 0.014967694878578186, + 0.015927499160170555, + 0.01411009207367897, + 0.026696499437093735, + 0.007278424222022295, + 0.0067975460551679134, + 0.011053099296987057, + 0.006101600825786591, + 0.007457571569830179, + 0.007079127244651318, + 0.007827497087419033, + 0.012513382360339165, + 0.0035507301799952984, + 0.0024929307401180267, + 0.004162859171628952, + 0.003108224831521511, + 0.003696332685649395, + 0.004725836683064699, + 0.005467214155942202, + 0.0074573722667992115, + 0.004914714023470879, + 0.0030828150920569897, + 0.004765297286212444, + 0.0032897989731281996, + 0.0037518099416047335, + 0.004687128122895956, + 0.005764417815953493, + 0.004137851297855377, + 0.04039440304040909, + 0.04075613617897034, + 0.02816435880959034, + 0.0456603467464447, + 0.03006705641746521, + 0.02549051120877266, + 0.04153983294963837, + 0.04316670447587967, + 0.006181362085044384, + 0.007701559457927942, + 0.009664228186011314, + 0.0063656349666416645, + 0.009778874926269054, + 0.006568532437086105, + 0.0053625633008778095, + 0.007250332273542881, + 0.006024675443768501, + 0.006240454502403736, + 0.005591203924268484, + 0.006491315085440874, + 0.009804613888263702, + 0.004430104047060013, + 0.00479532964527607, + 0.0058649745769798756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.016067780554294586 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810303658246994 + ], + "opposite_logprob_tokens": [ + -3.2881031036376953 + ], + "logprob_reward": -0.03810303658246994, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810303658246994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16192187368869781, + 0.03843441605567932, + 0.13097383081912994, + 0.09362314641475677, + 0.02143636718392372, + 0.028257843106985092, + 0.04192345216870308, + 0.1342717856168747, + 0.30356672406196594, + 0.04559050500392914 + ], + "credit_map": [ + 0.026122761890292168, + 0.014164279215037823, + 0.026465434581041336, + 0.008736833930015564, + 0.013447039760649204, + 0.0469822958111763, + 0.012070494703948498, + 0.013932746835052967, + 0.005256862845271826, + 0.002748975995928049, + 0.006349570117890835, + 0.003860610071569681, + 0.0033058328554034233, + 0.007610350381582975, + 0.004968512337654829, + 0.004333702847361565, + 0.017898811027407646, + 0.008039808832108974, + 0.01825546845793724, + 0.006095782853662968, + 0.00745089165866375, + 0.025167612358927727, + 0.01167682372033596, + 0.03638862818479538, + 0.014385468326508999, + 0.011026078835129738, + 0.015292071737349033, + 0.008564019575715065, + 0.009101900272071362, + 0.013027574867010117, + 0.012564306147396564, + 0.009661732241511345, + 0.0019753570668399334, + 0.002769005950540304, + 0.0031096593011170626, + 0.0018154759891331196, + 0.0030960352160036564, + 0.002606422873213887, + 0.00241976510733366, + 0.0036446459125727415, + 0.002682867692783475, + 0.0021424449514597654, + 0.0032093829941004515, + 0.0027184276841580868, + 0.0033223137725144625, + 0.003612235886976123, + 0.004556315019726753, + 0.00601385859772563, + 0.0054845549166202545, + 0.00909833051264286, + 0.004864771384745836, + 0.0023374503944069147, + 0.00860540010035038, + 0.004576368257403374, + 0.0031781301368027925, + 0.00377845112234354, + 0.015431939624249935, + 0.021999426186084747, + 0.015057321637868881, + 0.016547154635190964, + 0.01522613875567913, + 0.011430097743868828, + 0.018691327422857285, + 0.019888387992978096, + 0.033038437366485596, + 0.04359506815671921, + 0.03874659910798073, + 0.03844447806477547, + 0.05585520341992378, + 0.03137348219752312, + 0.028004182502627373, + 0.03450929746031761, + 0.0035218126140534878, + 0.005711887031793594, + 0.00619075121358037, + 0.004812012892216444, + 0.010116443037986755, + 0.005678308196365833, + 0.00463979784399271, + 0.0049194954335689545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.03810303658246994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, first-person close-up captures a clean hand. The tips of its thumb and index finger delicately grasp a few shrunken, yet vibrant red goji berries and several whole, dried chrysanthemums with pale yellow petals, each as delicate as paper. Soft afternoon sunlight streams through the side window, casting a warm golden glow over the cozy kitchen and illuminating the tiny dust particles floating in the air. The hand steadily and deliberately moves directly above a crystal-clear, round glass teapot. The sunlight reflects off the smooth, curved surface, casting dazzling, warm spots of light. The warm wooden countertop and a blurred kitchen corner in the background further enhance the homely atmosphere. As the fingertips gently release their grip, the red goji berries and pale yellow chrysanthemums drift gracefully in slow motion, finally scattering silently on the polished bottom of the pot. The entire action imbued with the tranquil ritual of preparing a cup of warming herbal tea, creating an atmosphere of ultimate relaxation, healing, and a sense of tranquility.", + "seed": 19002, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_14/index_7508_seed_19002_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12035586014389991, + "vlm_logprob": -0.12035586014389991, + "vlm_yes_no_margin": 3.0749999940395356, + "vlm_token_credit": -0.12035586014389991, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07088028639554977 + ], + "opposite_logprob_tokens": [ + -2.945880174636841 + ], + "logprob_reward": -0.07088028639554977, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07088028639554977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6658408045768738, + 0.08749201893806458, + 0.03665651008486748, + 0.06709551811218262, + 0.024573611095547676, + 0.015475341118872166, + 0.014300525188446045, + 0.0487523153424263, + 0.01659819297492504, + 0.0232151560485363 + ], + "credit_map": [ + 0.0960296094417572, + 0.1755274385213852, + 0.20400026440620422, + 0.020992932841181755, + 0.02792173996567726, + 0.06119764223694801, + 0.03647411987185478, + 0.0436970517039299, + 0.01241584774106741, + 0.021709436550736427, + 0.02649795077741146, + 0.0051552182994782925, + 0.003481403226032853, + 0.008667334914207458, + 0.005080976989120245, + 0.004483845550566912, + 0.004607431124895811, + 0.0034936561714857817, + 0.002779281698167324, + 0.004261041525751352, + 0.004769619088619947, + 0.005714386701583862, + 0.00524028018116951, + 0.0057908170856535435, + 0.006776751019060612, + 0.006159952376037836, + 0.007811904884874821, + 0.009512986056506634, + 0.011285645887255669, + 0.010375574231147766, + 0.006595042068511248, + 0.00857766717672348, + 0.0030883525032550097, + 0.0028087040409445763, + 0.0028867742512375116, + 0.002415598602965474, + 0.005236057098954916, + 0.00356718385592103, + 0.002758340910077095, + 0.001812600763514638, + 0.0018104070331901312, + 0.0010483755031600595, + 0.004313316661864519, + 0.001300011295825243, + 0.0015000206185504794, + 0.0009322307887487113, + 0.0027712928131222725, + 0.001799686928279698, + 0.002147649647668004, + 0.0016923986840993166, + 0.0021115299314260483, + 0.0017240502638742328, + 0.0016642861301079392, + 0.0018577729351818562, + 0.001571603468619287, + 0.0015312342438846827, + 0.0065231104381382465, + 0.009048997424542904, + 0.008292450569570065, + 0.006244153715670109, + 0.003533941460773349, + 0.00452270545065403, + 0.005799217149615288, + 0.004787737037986517, + 0.0021014921367168427, + 0.002085620304569602, + 0.0026161298155784607, + 0.001655703061260283, + 0.0012666629627346992, + 0.0014369826531037688, + 0.0035988399758934975, + 0.0018367613665759563, + 0.0022820676676928997, + 0.0028771585784852505, + 0.0025126358959823847, + 0.002757627284154296, + 0.0038300883024930954, + 0.0021367145236581564, + 0.003967057913541794, + 0.0028518070466816425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07088028639554977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012572435662150383 + ], + "opposite_logprob_tokens": [ + -4.387572288513184 + ], + "logprob_reward": -0.012572435662150383, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012572435662150383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6556841731071472, + 0.038751911371946335, + 0.024503014981746674, + 0.03692023828625679, + 0.03270003944635391, + 0.008033405058085918, + 0.022079268470406532, + 0.07972896099090576, + 0.04863559454679489, + 0.05296338349580765 + ], + "credit_map": [ + 0.09584349393844604, + 0.17361609637737274, + 0.19148750603199005, + 0.02022285759449005, + 0.03868281468749046, + 0.06140005216002464, + 0.03387375548481941, + 0.04055752232670784, + 0.007467256858944893, + 0.0112014040350914, + 0.006213757209479809, + 0.0019390330417081714, + 0.0032938227523118258, + 0.004233076237142086, + 0.0022899259347468615, + 0.002113630762323737, + 0.002684196690097451, + 0.0024301467929035425, + 0.0032841996289789677, + 0.003534609917551279, + 0.002613323274999857, + 0.003306790953502059, + 0.003702670568600297, + 0.0029470776207745075, + 0.002851436845958233, + 0.0021579861640930176, + 0.003639678005129099, + 0.002752197440713644, + 0.00800530705600977, + 0.011617086827754974, + 0.003734233323484659, + 0.002162311226129532, + 0.0035195262171328068, + 0.0026352994609624147, + 0.002292175777256489, + 0.0029846313409507275, + 0.006445517763495445, + 0.0035016080364584923, + 0.007788324728608131, + 0.003532951697707176, + 0.0009369529434479773, + 0.0006810606573708355, + 0.0012059265282005072, + 0.0006407895125448704, + 0.0010586421703919768, + 0.000784032978117466, + 0.0018966594943776727, + 0.0008293398423120379, + 0.005068739876151085, + 0.0062105245888233185, + 0.002076589036732912, + 0.001675045001320541, + 0.0021730302833020687, + 0.0015361977275460958, + 0.0016460512997582555, + 0.0016930887941271067, + 0.013694368302822113, + 0.01857321336865425, + 0.009629309177398682, + 0.00940164364874363, + 0.006102487910538912, + 0.006758293602615595, + 0.01035036239773035, + 0.005219274200499058, + 0.008072379976511002, + 0.009017471224069595, + 0.005728856660425663, + 0.00519277760758996, + 0.003575404640287161, + 0.00523021025583148, + 0.007452438119798899, + 0.004366051871329546, + 0.006481638643890619, + 0.006210747174918652, + 0.005289487075060606, + 0.005408445838838816, + 0.009533197619020939, + 0.004885207396000624, + 0.009515038691461086, + 0.0056396168656647205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.012572435662150383 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592818036675453 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592818036675453, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592818036675453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6429641246795654, + 0.12026460468769073, + 0.02821618877351284, + 0.04270849749445915, + 0.047424864023923874, + 0.02276201918721199, + 0.02804131805896759, + 0.019368186593055725, + 0.014461514540016651, + 0.03378863260149956 + ], + "credit_map": [ + 0.0938129872083664, + 0.17160247266292572, + 0.1970987319946289, + 0.01958496868610382, + 0.025980953127145767, + 0.059662118554115295, + 0.034808967262506485, + 0.04041297733783722, + 0.019751664251089096, + 0.032671574503183365, + 0.03264157474040985, + 0.006408222950994968, + 0.004767683334648609, + 0.01010957546532154, + 0.007353378925472498, + 0.006560934707522392, + 0.003146809758618474, + 0.002218451350927353, + 0.003380588721483946, + 0.0034118001349270344, + 0.003415563376620412, + 0.005008802283555269, + 0.0045171454548835754, + 0.0031170276924967766, + 0.0029856099281460047, + 0.0017731762491166592, + 0.0027092585805803537, + 0.0028796615079045296, + 0.01077278796583414, + 0.015056601725518703, + 0.002132665365934372, + 0.0043987371027469635, + 0.0053495559841394424, + 0.004985665902495384, + 0.004457581322640181, + 0.004080185201019049, + 0.0078005436807870865, + 0.003629056503996253, + 0.012815228663384914, + 0.004307046998292208, + 0.0031558857299387455, + 0.0025860448367893696, + 0.0029715728014707565, + 0.0022705893497914076, + 0.002098350552842021, + 0.0014427611604332924, + 0.005862170364707708, + 0.002374644624069333, + 0.00368545507080853, + 0.0028608888387680054, + 0.00360667472705245, + 0.0038407761603593826, + 0.003634239314123988, + 0.00301959621720016, + 0.003953668288886547, + 0.0034400199074298143, + 0.004288992378860712, + 0.0031385826878249645, + 0.002741995034739375, + 0.0022211077157408, + 0.00136879354249686, + 0.0016380969900637865, + 0.0020565176382660866, + 0.0019141003722324967, + 0.00245487317442894, + 0.0023708962835371494, + 0.001791056478396058, + 0.0015892658848315477, + 0.0009154254221357405, + 0.002081469399854541, + 0.0019157588249072433, + 0.0013427681988105178, + 0.0034880549646914005, + 0.0050002639181911945, + 0.0040536122396588326, + 0.0035267698112875223, + 0.005998013075441122, + 0.003183710156008601, + 0.004667700733989477, + 0.003870506538078189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.020592818036675453 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330007031559944 + ], + "opposite_logprob_tokens": [ + -3.7733001708984375 + ], + "logprob_reward": -0.02330007031559944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330007031559944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1939820945262909, + 0.1386609524488449, + 0.04002998769283295, + 0.17336247861385345, + 0.19757257401943207, + 0.028468241915106773, + 0.07951539009809494, + 0.08836794644594193, + 0.03711450845003128, + 0.022925863042473793 + ], + "credit_map": [ + 0.025742240250110626, + 0.03782579302787781, + 0.046500418335199356, + 0.010478993877768517, + 0.028534376993775368, + 0.017649400979280472, + 0.01594376191496849, + 0.011307117529213428, + 0.02411927655339241, + 0.0392509400844574, + 0.03532437980175018, + 0.007595903240144253, + 0.006091397255659103, + 0.011201268993318081, + 0.007842124439775944, + 0.007235666736960411, + 0.00592124043032527, + 0.005129518453031778, + 0.004594006575644016, + 0.004951775539666414, + 0.004728555213660002, + 0.007354713976383209, + 0.003949612379074097, + 0.003400564892217517, + 0.01967657171189785, + 0.009291760623455048, + 0.011375866830348969, + 0.009251758456230164, + 0.034393489360809326, + 0.056506820023059845, + 0.02135104313492775, + 0.011515173129737377, + 0.02135816216468811, + 0.013158154673874378, + 0.013339690864086151, + 0.01753576099872589, + 0.04280984774231911, + 0.016969310119748116, + 0.05229093134403229, + 0.020110707730054855, + 0.003391281934455037, + 0.0017330916598439217, + 0.007769002579152584, + 0.0020536822266876698, + 0.003595373360440135, + 0.0022879771422594786, + 0.004765528719872236, + 0.002872302196919918, + 0.012223560363054276, + 0.018711375072598457, + 0.009061010554432869, + 0.006318873725831509, + 0.008929884992539883, + 0.008748970925807953, + 0.00859781913459301, + 0.006923903711140156, + 0.014139602892100811, + 0.015984702855348587, + 0.015349200926721096, + 0.011349477805197239, + 0.006214400287717581, + 0.007974820211529732, + 0.009879535995423794, + 0.007476209197193384, + 0.005877481307834387, + 0.0044569289311766624, + 0.005740814842283726, + 0.00487090228125453, + 0.002151191933080554, + 0.004505976103246212, + 0.005403207615017891, + 0.004108003806322813, + 0.003112931502982974, + 0.002413524081930518, + 0.0036090114153921604, + 0.003278119722381234, + 0.001732655568048358, + 0.0021861731074750423, + 0.003822213737294078, + 0.0027712334413081408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.02330007031559944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47443369030952454 + ], + "opposite_logprob_tokens": [ + -0.9744336605072021 + ], + "logprob_reward": -0.47443369030952454, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.47443369030952454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6411828398704529, + 0.04081977903842926, + 0.011703718453645706, + 0.0315648652613163, + 0.1250600814819336, + 0.01573716104030609, + 0.04684433713555336, + 0.041665252298116684, + 0.024293487891554832, + 0.021128511056303978 + ], + "credit_map": [ + 0.09571515768766403, + 0.17273898422718048, + 0.1936156153678894, + 0.019793115556240082, + 0.027631046250462532, + 0.05829727649688721, + 0.03473883494734764, + 0.03865279257297516, + 0.006678554695099592, + 0.010584221221506596, + 0.00959128886461258, + 0.003124636597931385, + 0.002182371448725462, + 0.003513352945446968, + 0.0025449302047491074, + 0.0026004272513091564, + 0.0016876759473234415, + 0.0011275180149823427, + 0.0013046461390331388, + 0.002065790118649602, + 0.0013027731329202652, + 0.001633134321309626, + 0.0014476850628852844, + 0.0011344959493726492, + 0.004085018299520016, + 0.0029523305129259825, + 0.0030838772654533386, + 0.003092721803113818, + 0.005557764787226915, + 0.0069809271954, + 0.0028221704997122288, + 0.0029900555964559317, + 0.01329321414232254, + 0.009914633817970753, + 0.009169756434857845, + 0.014083431102335453, + 0.029515286907553673, + 0.013361613266170025, + 0.02089909464120865, + 0.014823047444224358, + 0.001379240071401, + 0.0011847087880596519, + 0.003745889523997903, + 0.0013641045661643147, + 0.002285293536260724, + 0.0011473868507891893, + 0.003099540015682578, + 0.0015309985028579831, + 0.008875444531440735, + 0.013132641091942787, + 0.004331967793405056, + 0.004050449933856726, + 0.004176728427410126, + 0.0045713381841778755, + 0.003887327155098319, + 0.003818443976342678, + 0.008361516520380974, + 0.009454537183046341, + 0.003849351080134511, + 0.004204487893730402, + 0.0023691202513873577, + 0.003190085757523775, + 0.006507650949060917, + 0.003728503594174981, + 0.0036462321877479553, + 0.0037489901296794415, + 0.004021123517304659, + 0.0032735413406044245, + 0.001820168225094676, + 0.00210880720987916, + 0.0030231119599193335, + 0.0026515123900026083, + 0.0024684867821633816, + 0.0029969795141369104, + 0.002834922634065151, + 0.002319671679288149, + 0.0028883852064609528, + 0.0019793743267655373, + 0.0025285568553954363, + 0.0031121333595365286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.47443369030952454 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, first-person close-up captures a clean hand. The tips of its thumb and index finger delicately grasp a few shrunken, yet vibrant red goji berries and several whole, dried chrysanthemums with pale yellow petals, each as delicate as paper. Soft afternoon sunlight streams through the side window, casting a warm golden glow over the cozy kitchen and illuminating the tiny dust particles floating in the air. The hand steadily and deliberately moves directly above a crystal-clear, round glass teapot. The sunlight reflects off the smooth, curved surface, casting dazzling, warm spots of light. The warm wooden countertop and a blurred kitchen corner in the background further enhance the homely atmosphere. As the fingertips gently release their grip, the red goji berries and pale yellow chrysanthemums drift gracefully in slow motion, finally scattering silently on the polished bottom of the pot. The entire action imbued with the tranquil ritual of preparing a cup of warming herbal tea, creating an atmosphere of ultimate relaxation, healing, and a sense of tranquility.", + "seed": 19003, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_14/index_7508_seed_19003_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02717865165323019, + "vlm_logprob": -0.02717865165323019, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02717865165323019, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030377164483070374 + ], + "opposite_logprob_tokens": [ + -3.905377149581909 + ], + "logprob_reward": -0.030377164483070374, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.030377164483070374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10012129694223404, + 0.14211654663085938, + 0.1620929092168808, + 0.10877075791358948, + 0.1902448683977127, + 0.049616456031799316, + 0.0608142651617527, + 0.09402664750814438, + 0.06944814324378967, + 0.022748159244656563 + ], + "credit_map": [ + 0.016956482082605362, + 0.013266733847558498, + 0.011509130708873272, + 0.014213415794074535, + 0.013762649148702621, + 0.010583216324448586, + 0.009763164445757866, + 0.010066496208310127, + 0.037379927933216095, + 0.020763637498021126, + 0.010955994948744774, + 0.01697125844657421, + 0.013496874831616879, + 0.017693523317575455, + 0.01310522761195898, + 0.011750081554055214, + 0.025758450850844383, + 0.042388372123241425, + 0.021123571321368217, + 0.013369385153055191, + 0.017290860414505005, + 0.01628863252699375, + 0.01255317684262991, + 0.013320454396307468, + 0.009256664663553238, + 0.011992020532488823, + 0.014494876377284527, + 0.017401762306690216, + 0.013816854916512966, + 0.014882980845868587, + 0.013176622800529003, + 0.013748973608016968, + 0.012102615088224411, + 0.016107188537716866, + 0.031227467581629753, + 0.018360942602157593, + 0.046277593821287155, + 0.03706079348921776, + 0.01498471014201641, + 0.014123531058430672, + 0.004134725779294968, + 0.005619775969535112, + 0.006317178253084421, + 0.00596635090187192, + 0.010063782334327698, + 0.0068567367270588875, + 0.006023031659424305, + 0.004634869284927845, + 0.006251646671444178, + 0.005814943928271532, + 0.01196087896823883, + 0.007046537473797798, + 0.006716618314385414, + 0.006926944013684988, + 0.00865186844021082, + 0.007444814778864384, + 0.011090340092778206, + 0.0056648459285497665, + 0.023305250331759453, + 0.014280779287219048, + 0.004820467904210091, + 0.008537461049854755, + 0.011062136851251125, + 0.015265356749296188, + 0.0076005542650818825, + 0.009569133631885052, + 0.01198411826044321, + 0.00687019620090723, + 0.007526957429945469, + 0.0076149385422468185, + 0.010097194463014603, + 0.008185039274394512, + 0.0036978067364543676, + 0.002607678296044469, + 0.0034666284918785095, + 0.0028427790384739637, + 0.0017500835238024592, + 0.00223740772344172, + 0.0029294986743479967, + 0.0032162743154913187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.030377164483070374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004654288291931152 + ], + "opposite_logprob_tokens": [ + -5.379654407501221 + ], + "logprob_reward": -0.004654288291931152, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004654288291931152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14736661314964294, + 0.07527080178260803, + 0.18132048845291138, + 0.06954371929168701, + 0.1734943836927414, + 0.05554008111357689, + 0.10974515229463577, + 0.11369059234857559, + 0.04512646421790123, + 0.028901688754558563 + ], + "credit_map": [ + 0.014321509748697281, + 0.01864287443459034, + 0.014699984341859818, + 0.018053509294986725, + 0.0259189922362566, + 0.025051919743418694, + 0.0173789169639349, + 0.013298915699124336, + 0.015911411494016647, + 0.011468973942101002, + 0.005093463230878115, + 0.009867066517472267, + 0.006957894656807184, + 0.00848749466240406, + 0.011089695617556572, + 0.006394806317985058, + 0.025398647412657738, + 0.018823472782969475, + 0.02057756297290325, + 0.021966608241200447, + 0.023303519934415817, + 0.02344844490289688, + 0.028064070269465446, + 0.01973816566169262, + 0.006387729197740555, + 0.006741637364029884, + 0.009126548655331135, + 0.012399447150528431, + 0.006658170372247696, + 0.008960851468145847, + 0.00860059354454279, + 0.0106687406077981, + 0.011616040952503681, + 0.015888916328549385, + 0.02712571993470192, + 0.013887315057218075, + 0.04547140374779701, + 0.03472210839390755, + 0.013084960170090199, + 0.011697916314005852, + 0.0038131719920784235, + 0.005169373005628586, + 0.006029256619513035, + 0.0080404719337821, + 0.00821540504693985, + 0.009899323806166649, + 0.007766332942992449, + 0.0066067492589354515, + 0.009408666752278805, + 0.01231156475841999, + 0.023248722776770592, + 0.010049539618194103, + 0.014775851741433144, + 0.014471014030277729, + 0.011149330995976925, + 0.014330465346574783, + 0.024597978219389915, + 0.011636127717792988, + 0.03763479366898537, + 0.004744417499750853, + 0.01410150621086359, + 0.009777795523405075, + 0.005984188988804817, + 0.005213786382228136, + 0.004751990549266338, + 0.004773527383804321, + 0.008292367681860924, + 0.006562347523868084, + 0.00463187787681818, + 0.0038668785709887743, + 0.005902995355427265, + 0.006344479508697987, + 0.00398907158523798, + 0.004056923557072878, + 0.0036469290498644114, + 0.004592619836330414, + 0.0016438078600913286, + 0.0022062486968934536, + 0.0034362333826720715, + 0.005329854786396027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.004654288291931152 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419929601252079 + ], + "opposite_logprob_tokens": [ + -4.264199256896973 + ], + "logprob_reward": -0.01419929601252079, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419929601252079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18047785758972168, + 0.08032657951116562, + 0.06504685431718826, + 0.09500829875469208, + 0.12232717871665955, + 0.03716881945729256, + 0.11767472326755524, + 0.16579104959964752, + 0.0969553291797638, + 0.039223261177539825 + ], + "credit_map": [ + 0.021437885239720345, + 0.02404867671430111, + 0.017831997945904732, + 0.022285187616944313, + 0.028998397290706635, + 0.022419916465878487, + 0.022822057828307152, + 0.0206337571144104, + 0.009180799126625061, + 0.008146783336997032, + 0.00863445084542036, + 0.012574671767652035, + 0.011278144083917141, + 0.010949677787721157, + 0.012179721146821976, + 0.007382339797914028, + 0.010129004716873169, + 0.010193919762969017, + 0.00733556505292654, + 0.009356147609651089, + 0.005847690161317587, + 0.005636901129037142, + 0.009351002983748913, + 0.007196628954261541, + 0.010890182107686996, + 0.00984435435384512, + 0.012300941161811352, + 0.01859315298497677, + 0.011060519143939018, + 0.008246940560638905, + 0.010621467605233192, + 0.013450748287141323, + 0.010304718278348446, + 0.016264550387859344, + 0.018212471157312393, + 0.01230397168546915, + 0.02253427542746067, + 0.020790260285139084, + 0.012706148438155651, + 0.009210793301463127, + 0.0030921767465770245, + 0.003375814063474536, + 0.004566204268485308, + 0.0036088177002966404, + 0.006953501608222723, + 0.005663394927978516, + 0.00516607565805316, + 0.00474283704534173, + 0.01482595968991518, + 0.015419986099004745, + 0.022275591269135475, + 0.009283208288252354, + 0.021007904782891273, + 0.014456558972597122, + 0.009519532322883606, + 0.010885987430810928, + 0.03101622872054577, + 0.013368017971515656, + 0.06343550235033035, + 0.010678205639123917, + 0.013188647106289864, + 0.016499413177371025, + 0.007996710948646069, + 0.009608332999050617, + 0.01327692810446024, + 0.018180372193455696, + 0.01371703576296568, + 0.01190586294978857, + 0.008161579258739948, + 0.007884337566792965, + 0.010428615845739841, + 0.013400599360466003, + 0.0065581887029111385, + 0.005303835961967707, + 0.005163345951586962, + 0.00588771840557456, + 0.0028381627053022385, + 0.004095016047358513, + 0.004636874422430992, + 0.004740124102681875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.01419929601252079 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628121413290501 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007628121413290501, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628121413290501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09733027219772339, + 0.08625621348619461, + 0.04157225787639618, + 0.11736863106489182, + 0.2686424255371094, + 0.033361464738845825, + 0.07686726748943329, + 0.16004659235477448, + 0.08571615070104599, + 0.032838743180036545 + ], + "credit_map": [ + 0.011458637192845345, + 0.014284001663327217, + 0.010332176461815834, + 0.011337637901306152, + 0.017606906592845917, + 0.014906639233231544, + 0.00848171766847372, + 0.008922549895942211, + 0.01459516305476427, + 0.012096825987100601, + 0.008937746286392212, + 0.012368773110210896, + 0.008573172613978386, + 0.011377144604921341, + 0.01043915655463934, + 0.007868234068155289, + 0.006789418403059244, + 0.0072095682844519615, + 0.004432522226125002, + 0.004199118819087744, + 0.004789865110069513, + 0.004619603510946035, + 0.0049230968579649925, + 0.004609061870723963, + 0.011429406702518463, + 0.013145615346729755, + 0.015394041314721107, + 0.021379994228482246, + 0.012945025227963924, + 0.014884351752698421, + 0.012788994237780571, + 0.015401204116642475, + 0.017881246283650398, + 0.02455076016485691, + 0.044247519224882126, + 0.017898911610245705, + 0.07186397165060043, + 0.05404743552207947, + 0.01861572079360485, + 0.019536864012479782, + 0.003299351315945387, + 0.0035194484516978264, + 0.004414914641529322, + 0.00475583178922534, + 0.006630966905504465, + 0.004116798285394907, + 0.00395470205694437, + 0.002669450594112277, + 0.0075796316377818584, + 0.007076011970639229, + 0.008422531187534332, + 0.00978653971105814, + 0.0075931549072265625, + 0.010435526259243488, + 0.012888263911008835, + 0.013085608370602131, + 0.014536917209625244, + 0.012121033854782581, + 0.035252396017313004, + 0.028137294575572014, + 0.007719710469245911, + 0.013436547480523586, + 0.022904828190803528, + 0.025937864556908607, + 0.010398879647254944, + 0.009777083061635494, + 0.018073679879307747, + 0.012238522991538048, + 0.009485295042395592, + 0.006586362142115831, + 0.009766520000994205, + 0.009389806538820267, + 0.004845789168030024, + 0.004536904394626617, + 0.003919267561286688, + 0.0046865190379321575, + 0.001758504076860845, + 0.003432883182540536, + 0.004604157991707325, + 0.005054717417806387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.007628121413290501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903438806533813 + ], + "opposite_logprob_tokens": [ + -2.5790343284606934 + ], + "logprob_reward": -0.07903438806533813, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903438806533813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3398815393447876, + 0.17920897901058197, + 0.031540390104055405, + 0.15118348598480225, + 0.03789212927222252, + 0.014586414210498333, + 0.03871883079409599, + 0.09057317674160004, + 0.06152326613664627, + 0.054891813546419144 + ], + "credit_map": [ + 0.05681752786040306, + 0.049891021102666855, + 0.05111636966466904, + 0.04415983334183693, + 0.04678259789943695, + 0.03411378338932991, + 0.026573294773697853, + 0.030427111312747, + 0.04428601264953613, + 0.028644781559705734, + 0.014760964550077915, + 0.019983427599072456, + 0.01533396728336811, + 0.02234838157892227, + 0.01884523779153824, + 0.015006205067038536, + 0.004479177296161652, + 0.004826288670301437, + 0.00417554285377264, + 0.004000174812972546, + 0.0032004646491259336, + 0.0037528162356466055, + 0.003702664515003562, + 0.003403259674087167, + 0.016564365476369858, + 0.02691642940044403, + 0.017517732456326485, + 0.022875167429447174, + 0.018420033156871796, + 0.01377284899353981, + 0.019005462527275085, + 0.01611143909394741, + 0.0039993091486394405, + 0.00472271116450429, + 0.005091921892017126, + 0.002939122961834073, + 0.0066918907687067986, + 0.006106378044933081, + 0.004499448463320732, + 0.00384134822525084, + 0.0019718275871127844, + 0.0015390720218420029, + 0.0020446705166250467, + 0.0016896326560527086, + 0.002016465412452817, + 0.0016538465861231089, + 0.002267556032165885, + 0.0014033425832167268, + 0.0033856325317174196, + 0.004146311432123184, + 0.00684052100405097, + 0.0036828815937042236, + 0.004872462246567011, + 0.0049875094555318356, + 0.005761512089520693, + 0.005042002070695162, + 0.018916333094239235, + 0.006461894605308771, + 0.03460586816072464, + 0.005919787101447582, + 0.006423125043511391, + 0.008525694720447063, + 0.004394717048853636, + 0.00532575836405158, + 0.010432779788970947, + 0.013942110352218151, + 0.00882315170019865, + 0.005710210185497999, + 0.006125138606876135, + 0.005037487018853426, + 0.005800558719784021, + 0.005651829764246941, + 0.007471328601241112, + 0.0076244426891207695, + 0.00836230255663395, + 0.008878248743712902, + 0.0031889118254184723, + 0.004764639772474766, + 0.006673512049019337, + 0.007928428240120411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07903438806533813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, first-person close-up captures a clean hand. The tips of its thumb and index finger delicately grasp a few shrunken, yet vibrant red goji berries and several whole, dried chrysanthemums with pale yellow petals, each as delicate as paper. Soft afternoon sunlight streams through the side window, casting a warm golden glow over the cozy kitchen and illuminating the tiny dust particles floating in the air. The hand steadily and deliberately moves directly above a crystal-clear, round glass teapot. The sunlight reflects off the smooth, curved surface, casting dazzling, warm spots of light. The warm wooden countertop and a blurred kitchen corner in the background further enhance the homely atmosphere. As the fingertips gently release their grip, the red goji berries and pale yellow chrysanthemums drift gracefully in slow motion, finally scattering silently on the polished bottom of the pot. The entire action imbued with the tranquil ritual of preparing a cup of warming herbal tea, creating an atmosphere of ultimate relaxation, healing, and a sense of tranquility.", + "seed": 19004, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_14/index_7508_seed_19004_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1318723970092833, + "vlm_logprob": -0.1318723970092833, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.1318723970092833, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09471571445465088 + ], + "opposite_logprob_tokens": [ + -2.5947155952453613 + ], + "logprob_reward": -0.09471571445465088, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09471571445465088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15000402927398682, + 0.052584923803806305, + 0.07951264083385468, + 0.04820751026272774, + 0.14380577206611633, + 0.04527457430958748, + 0.2992664873600006, + 0.05441184341907501, + 0.05929278954863548, + 0.06763942539691925 + ], + "credit_map": [ + 0.01529635675251484, + 0.011770264245569706, + 0.017407631501555443, + 0.0149918207898736, + 0.024703945964574814, + 0.02311917208135128, + 0.014630393125116825, + 0.028084464371204376, + 0.008730410598218441, + 0.003176165744662285, + 0.007094495464116335, + 0.0037022540345788, + 0.004584912676364183, + 0.005810260307043791, + 0.010686676017940044, + 0.008799746632575989, + 0.014118830673396587, + 0.004430735949426889, + 0.00657729571685195, + 0.004787155892699957, + 0.007541576866060495, + 0.008865803480148315, + 0.021093998104333878, + 0.012097234837710857, + 0.010675575584173203, + 0.0028605798725038767, + 0.003450562711805105, + 0.003077712608501315, + 0.002811047015711665, + 0.006066884379833937, + 0.006335792131721973, + 0.012929356656968594, + 0.01900782249867916, + 0.010038728825747967, + 0.015978168696165085, + 0.011001222766935825, + 0.012479607947170734, + 0.012988747097551823, + 0.02198093757033348, + 0.04033054783940315, + 0.00793168693780899, + 0.004733108449727297, + 0.003956549800932407, + 0.003957842011004686, + 0.007058848161250353, + 0.0038751130923628807, + 0.006703170482069254, + 0.007058253511786461, + 0.04595063626766205, + 0.0090754684060812, + 0.008172381669282913, + 0.007705578114837408, + 0.058898162096738815, + 0.011353490874171257, + 0.1325148344039917, + 0.025595935061573982, + 0.005140781868249178, + 0.004965381231158972, + 0.0035134076606482267, + 0.00431173387914896, + 0.00725106755271554, + 0.005087071098387241, + 0.010602729395031929, + 0.013539671897888184, + 0.004911098629236221, + 0.010310027748346329, + 0.003571345703676343, + 0.003281135344877839, + 0.014201918616890907, + 0.004276054911315441, + 0.010428601875901222, + 0.008312607184052467, + 0.007630266714841127, + 0.01380726508796215, + 0.003646271303296089, + 0.006060032639652491, + 0.009989380836486816, + 0.004879250191152096, + 0.009240158833563328, + 0.012386800721287727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.09471571445465088 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006772658322006464 + ], + "opposite_logprob_tokens": [ + -5.006772518157959 + ], + "logprob_reward": -0.006772658322006464, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006772658322006464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07508909702301025, + 0.06107529625296593, + 0.14331671595573425, + 0.16264179348945618, + 0.09768526256084442, + 0.03158927336335182, + 0.2173680067062378, + 0.0483434833586216, + 0.07543742656707764, + 0.087453693151474 + ], + "credit_map": [ + 0.01003611646592617, + 0.007807161659002304, + 0.010338200256228447, + 0.006341671105474234, + 0.012822188436985016, + 0.011555146425962448, + 0.00562006002292037, + 0.010568543337285519, + 0.00683525949716568, + 0.004043759312480688, + 0.007001413498073816, + 0.00454862741753459, + 0.013791394419968128, + 0.01001102477312088, + 0.00834435224533081, + 0.006499461829662323, + 0.014613274484872818, + 0.01118730939924717, + 0.012064380571246147, + 0.007819101214408875, + 0.018808206543326378, + 0.02395205944776535, + 0.03231168910861015, + 0.02256067655980587, + 0.03362525999546051, + 0.009763096459209919, + 0.008910460397601128, + 0.007456568535417318, + 0.013462580740451813, + 0.0211003627628088, + 0.02143690176308155, + 0.046886544674634933, + 0.014677947387099266, + 0.0053604645654559135, + 0.01006334275007248, + 0.00729815848171711, + 0.005752121098339558, + 0.008395434357225895, + 0.014940696768462658, + 0.031197089701890945, + 0.0034141477663069963, + 0.0028855556156486273, + 0.003393368562683463, + 0.0027018676046282053, + 0.004809848964214325, + 0.003338377457112074, + 0.005711324978619814, + 0.005334779620170593, + 0.03625478968024254, + 0.008138486184179783, + 0.005920985713601112, + 0.005557456985116005, + 0.03860123082995415, + 0.008510176092386246, + 0.0964922234416008, + 0.01789264939725399, + 0.008118879050016403, + 0.00852470938116312, + 0.002136490773409605, + 0.0036817302461713552, + 0.00629481952637434, + 0.003708440810441971, + 0.006606184411793947, + 0.009272227063775063, + 0.006401499267667532, + 0.014782432466745377, + 0.0040644872933626175, + 0.005838950630277395, + 0.017502252012491226, + 0.005980513524264097, + 0.010165615938603878, + 0.010701671242713928, + 0.0051773060113191605, + 0.009814721532166004, + 0.0024807353038340807, + 0.00537226814776659, + 0.0069987052120268345, + 0.004010540433228016, + 0.02914913184940815, + 0.024450283497571945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.006772658322006464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191752955317497 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.018191752955317497, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191752955317497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1223575547337532, + 0.061288002878427505, + 0.08712642639875412, + 0.10904484242200851, + 0.07051926106214523, + 0.10818469524383545, + 0.26309943199157715, + 0.0600789338350296, + 0.06077656149864197, + 0.05752434954047203 + ], + "credit_map": [ + 0.01559193804860115, + 0.00866174791008234, + 0.015987997874617577, + 0.015508521348237991, + 0.014158300124108791, + 0.014030637219548225, + 0.01479834970086813, + 0.023620061576366425, + 0.008897962979972363, + 0.003920127637684345, + 0.008012033998966217, + 0.006489090155810118, + 0.004783837124705315, + 0.004444822669029236, + 0.014369698241353035, + 0.010370435193181038, + 0.014437914825975895, + 0.004870068281888962, + 0.008945980109274387, + 0.006402478087693453, + 0.006345665082335472, + 0.0080013582482934, + 0.02221345715224743, + 0.015909506008028984, + 0.011586637236177921, + 0.008430112153291702, + 0.012633681297302246, + 0.009469379670917988, + 0.0164497047662735, + 0.01991410367190838, + 0.017691390588879585, + 0.01286984235048294, + 0.011589741334319115, + 0.0076070851646363735, + 0.008561007678508759, + 0.004850823897868395, + 0.00904388539493084, + 0.007246428169310093, + 0.008322703652083874, + 0.013297582045197487, + 0.015097768045961857, + 0.01052484568208456, + 0.010955053381621838, + 0.007185046561062336, + 0.01581747829914093, + 0.011538757011294365, + 0.019811047241091728, + 0.017254697158932686, + 0.04430047422647476, + 0.010146251879632473, + 0.012728041969239712, + 0.011265762150287628, + 0.04132545366883278, + 0.01139301247894764, + 0.10362469404935837, + 0.028315745294094086, + 0.007123294752091169, + 0.00928223691880703, + 0.0037370265927165747, + 0.0068482100032269955, + 0.009736698120832443, + 0.005675080697983503, + 0.007804003544151783, + 0.009872383438050747, + 0.007996578700840473, + 0.008190082386136055, + 0.0064795780926942825, + 0.0054802908562123775, + 0.007499543018639088, + 0.004802893381565809, + 0.010836915113031864, + 0.009490675292909145, + 0.0045241545885801315, + 0.006197414360940456, + 0.0034773361403495073, + 0.00511913001537323, + 0.0066472673788666725, + 0.0033391264732927084, + 0.012699801474809647, + 0.015520119108259678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.018191752955317497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607704907655716 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.01607704907655716, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607704907655716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12835480272769928, + 0.06573959439992905, + 0.07570794224739075, + 0.12429312616586685, + 0.13780489563941956, + 0.08218546956777573, + 0.09718763828277588, + 0.10974330455064774, + 0.042577967047691345, + 0.13640530407428741 + ], + "credit_map": [ + 0.012086912989616394, + 0.009117137640714645, + 0.01514280866831541, + 0.009511729702353477, + 0.016377205029129982, + 0.01976316049695015, + 0.020824234932661057, + 0.02553161047399044, + 0.013991326093673706, + 0.0038002252113074064, + 0.006101727951318026, + 0.004055715166032314, + 0.0057007260620594025, + 0.004900429397821426, + 0.019209951162338257, + 0.007979482412338257, + 0.00846057292073965, + 0.005454347003251314, + 0.010774505324661732, + 0.0078111919574439526, + 0.0054307663813233376, + 0.007739482447504997, + 0.012804584577679634, + 0.017232485115528107, + 0.026698537170886993, + 0.008078167214989662, + 0.0067482683807611465, + 0.005065887700766325, + 0.01325254701077938, + 0.015543268993496895, + 0.01475459337234497, + 0.034151848405599594, + 0.02059847116470337, + 0.008442536927759647, + 0.01371721364557743, + 0.010541599243879318, + 0.012641534209251404, + 0.01413024589419365, + 0.019277192652225494, + 0.038456086069345474, + 0.014059387147426605, + 0.005956556648015976, + 0.0075403549708426, + 0.009854177013039589, + 0.009884217754006386, + 0.008378459140658379, + 0.0120572904124856, + 0.014455020427703857, + 0.010716190561652184, + 0.009414120577275753, + 0.00885594729334116, + 0.007299371529370546, + 0.019475119188427925, + 0.009108219295740128, + 0.020429911091923714, + 0.011888752691447735, + 0.009075124748051167, + 0.00873484369367361, + 0.003865820588544011, + 0.005578839220106602, + 0.011965160258114338, + 0.003979570232331753, + 0.041103921830654144, + 0.025440014898777008, + 0.0032000120263546705, + 0.006623722147196531, + 0.0020005609840154648, + 0.004723175894469023, + 0.008784754201769829, + 0.003274323884397745, + 0.0071634566411376, + 0.006807955447584391, + 0.009495697915554047, + 0.010447990149259567, + 0.004080840852111578, + 0.0066937715746462345, + 0.009445472620427608, + 0.0047758533619344234, + 0.04661745950579643, + 0.04484819248318672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.01607704907655716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a steady first-person close-up of a clean hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are shrunken red goji berries and whole dried chrysanthemums with pale yellow petals visible in the hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the hand move steadily above a crystal-clear round glass teapot before releasing the items?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the goji berries and chrysanthemums drift in slow motion and scatter onto the bottom of the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft afternoon sunlight streaming through a side window, illuminating dust particles in a cozy kitchen?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5236048102378845 + ], + "opposite_logprob_tokens": [ + -0.8986048102378845 + ], + "logprob_reward": -0.5236048102378845, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5236048102378845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2828455865383148, + 0.04051060229539871, + 0.06876784563064575, + 0.09504098445177078, + 0.0478908009827137, + 0.10750889778137207, + 0.06974683701992035, + 0.1217670813202858, + 0.06213545799255371, + 0.10378588736057281 + ], + "credit_map": [ + 0.03722844645380974, + 0.027391504496335983, + 0.03630121797323227, + 0.04047520458698273, + 0.02470884658396244, + 0.02678096480667591, + 0.028871962800621986, + 0.06108744442462921, + 0.008037054911255836, + 0.0022530758287757635, + 0.003906976897269487, + 0.003868419909849763, + 0.0028328411281108856, + 0.0038701186422258615, + 0.008758284151554108, + 0.006983827333897352, + 0.010050302371382713, + 0.0051780990324914455, + 0.00710704131051898, + 0.003527899505570531, + 0.009097162634134293, + 0.011311385780572891, + 0.014220917597413063, + 0.008275032043457031, + 0.019674625247716904, + 0.005282304249703884, + 0.008407632820308208, + 0.004109214525669813, + 0.006959870923310518, + 0.010984636843204498, + 0.014196724630892277, + 0.025425972416996956, + 0.004671537317335606, + 0.005592640955001116, + 0.003329653525725007, + 0.0029191658832132816, + 0.00739431194961071, + 0.005556993652135134, + 0.008269209414720535, + 0.010157287120819092, + 0.01834229752421379, + 0.01008335780352354, + 0.00865256693214178, + 0.010205178521573544, + 0.016530754044651985, + 0.009741263464093208, + 0.01379274670034647, + 0.020160725340247154, + 0.010267852805554867, + 0.00616386067122221, + 0.007531231269240379, + 0.00530612375587225, + 0.014227514155209064, + 0.006589933298528194, + 0.007919354364275932, + 0.0117409722879529, + 0.019477158784866333, + 0.00931078102439642, + 0.003082823008298874, + 0.0056082624942064285, + 0.013142049312591553, + 0.004431334789842367, + 0.04969437047839165, + 0.017020300030708313, + 0.006973798852413893, + 0.008304290473461151, + 0.006320462562143803, + 0.0053787254728376865, + 0.009151068516075611, + 0.007180692162364721, + 0.010875140316784382, + 0.007951276376843452, + 0.012743765488266945, + 0.020981917157769203, + 0.004061535932123661, + 0.007418154738843441, + 0.01823807694017887, + 0.006849328987300396, + 0.014367913827300072, + 0.0191251989454031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.5236048102378845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_15/index_tensor([9549, 9549])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_15/index_tensor([9549, 9549])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fbfbc97f329585dee873de1ebfd688a82d0d2405 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_15/index_tensor([9549, 9549])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic wide-angle drone shot, beginning at a low angle and slowly zooming out, reveals a section of the ancient Great Wall bathed in dramatic dusk light, perched atop a steep, rugged ridge. Constructed of weathered gray boulders and mottled blue bricks, the wall's surface is covered in age-worn cracks and moss. The setting sun casts long, twisting shadows across its rough surface. Suddenly, accompanied by a low tremor, a majestic beacon tower structure begins to shake violently, sending tiny pieces of gravel and dust cascading down from the cracks between the bricks. Immediately afterwards, an unstoppable chain reaction broke out. The huge brick and stone structure disintegrated in an instant. Tons of bricks and stones exploded and poured down like a landslide. The collapsed structure turned into a huge, rolling cloud of yellow-brown dust. This cloud of dust was dyed a fiery orange-red under the reflection of the setting sun and quickly engulfed the hillside below. The whole picture was filled with an epic sense of destruction and the ruthless power of time, creating a grand and tragic atmosphere.", + "seed": 27770, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_15/index_9549_seed_27770_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.24559910940006374, + "vlm_logprob": -0.24559910940006374, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.24559910940006374, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7047558426856995 + ], + "opposite_logprob_tokens": [ + -0.7047558426856995 + ], + "logprob_reward": -0.7047558426856995, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7047558426856995, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13524064421653748, + 0.05916185304522514, + 0.036682792007923126, + 0.058968160301446915, + 0.09866024553775787, + 0.09153562039136887, + 0.1471341997385025, + 0.146861270070076, + 0.15338198840618134, + 0.07237321138381958 + ], + "credit_map": [ + 0.01208591740578413, + 0.0170688945800066, + 0.01652253419160843, + 0.019351910799741745, + 0.01599317044019699, + 0.021520638838410378, + 0.015814077109098434, + 0.016883498057723045, + 0.0047385068610310555, + 0.007154094520956278, + 0.008105389773845673, + 0.007684031501412392, + 0.006809587590396404, + 0.01068483479321003, + 0.006149278022348881, + 0.007836129516363144, + 0.003724024398252368, + 0.006249844096601009, + 0.0027207364328205585, + 0.003841735888272524, + 0.005364068318158388, + 0.0062858485616743565, + 0.004479625727981329, + 0.004016905557364225, + 0.006396380253136158, + 0.007308597210794687, + 0.009653765708208084, + 0.009969407692551613, + 0.00615431135520339, + 0.010093661025166512, + 0.004335156641900539, + 0.005056881811469793, + 0.014178084209561348, + 0.011219018138945103, + 0.010668370872735977, + 0.017675653100013733, + 0.012093741446733475, + 0.011576974764466286, + 0.008824645541608334, + 0.01242375373840332, + 0.009095560759305954, + 0.011913642287254333, + 0.00932955089956522, + 0.01384007465094328, + 0.009862051345407963, + 0.012657332234084606, + 0.01165904849767685, + 0.013178367167711258, + 0.01264189649373293, + 0.01499107014387846, + 0.05167050287127495, + 0.019719384610652924, + 0.008228977210819721, + 0.011066241189837456, + 0.011116798967123032, + 0.01769931986927986, + 0.018160071223974228, + 0.02651721052825451, + 0.019652923569083214, + 0.020547784864902496, + 0.012862890027463436, + 0.012068498879671097, + 0.011958062648773193, + 0.025093821808695793, + 0.022741198539733887, + 0.022985247895121574, + 0.03007531352341175, + 0.02616875059902668, + 0.014875379391014576, + 0.010471201501786709, + 0.012902618385851383, + 0.013162284158170223, + 0.006611732300370932, + 0.013259431347250938, + 0.009340394288301468, + 0.012317035347223282, + 0.009760010056197643, + 0.004849624820053577, + 0.008547660894691944, + 0.00768732326105237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.7047558426856995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813689410686493 + ], + "opposite_logprob_tokens": [ + -1.4063689708709717 + ], + "logprob_reward": -0.2813689410686493, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813689410686493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18269988894462585, + 0.1196088045835495, + 0.10011951625347137, + 0.08833880722522736, + 0.06655237823724747, + 0.07934554666280746, + 0.15261752903461456, + 0.10383865237236023, + 0.05752196162939072, + 0.04935690760612488 + ], + "credit_map": [ + 0.01240968331694603, + 0.015262920409440994, + 0.016909770667552948, + 0.02980518899857998, + 0.023531615734100342, + 0.03034728579223156, + 0.026521332561969757, + 0.027912093326449394, + 0.007102789357304573, + 0.00940767303109169, + 0.010422790423035622, + 0.016913417726755142, + 0.01622132398188114, + 0.02134830132126808, + 0.0144895538687706, + 0.023702960461378098, + 0.008576000109314919, + 0.01109645888209343, + 0.009632468223571777, + 0.012598768807947636, + 0.017227210104465485, + 0.0191110260784626, + 0.01064694207161665, + 0.011230646632611752, + 0.006874344311654568, + 0.007243082392960787, + 0.007758967112749815, + 0.00967266783118248, + 0.010793800465762615, + 0.01691710203886032, + 0.011075579561293125, + 0.018003258854150772, + 0.009311643429100513, + 0.008057128638029099, + 0.009251149371266365, + 0.01157708652317524, + 0.008280046284198761, + 0.007741311565041542, + 0.0059825885109603405, + 0.006351419724524021, + 0.004774013999849558, + 0.007144291885197163, + 0.009225748479366302, + 0.009203320369124413, + 0.008905893191695213, + 0.012881728820502758, + 0.015013511292636395, + 0.012197034433484077, + 0.015309187583625317, + 0.014273918233811855, + 0.0518389567732811, + 0.018035324290394783, + 0.0099323196336627, + 0.018239401280879974, + 0.012770771980285645, + 0.01221765112131834, + 0.009157105349004269, + 0.012404411099851131, + 0.013802885077893734, + 0.020128825679421425, + 0.014327941462397575, + 0.010425812564790249, + 0.009286923334002495, + 0.014304745942354202, + 0.006824010517448187, + 0.008991645649075508, + 0.009637748822569847, + 0.008678409270942211, + 0.008969677612185478, + 0.00505074393004179, + 0.004749929532408714, + 0.0046197944320738316, + 0.005275794304907322, + 0.007269266527146101, + 0.006742845289409161, + 0.0095660500228405, + 0.007111973129212856, + 0.0029571291524916887, + 0.005511678289622068, + 0.004922169726341963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.2813689410686493 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015765756368637 + ], + "opposite_logprob_tokens": [ + -1.701576590538025 + ], + "logprob_reward": -0.2015765756368637, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015765756368637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10197685658931732, + 0.025840193033218384, + 0.01554790697991848, + 0.037977200001478195, + 0.0720420628786087, + 0.07244706898927689, + 0.5759180188179016, + 0.025261398404836655, + 0.04745972529053688, + 0.02552957646548748 + ], + "credit_map": [ + 0.009021548554301262, + 0.010203945450484753, + 0.01608230546116829, + 0.01438052486628294, + 0.011575473472476006, + 0.017935404554009438, + 0.01248736958950758, + 0.01029027346521616, + 0.0021638558246195316, + 0.002245633164420724, + 0.004266747273504734, + 0.0036171316169202328, + 0.002659629564732313, + 0.004802619572728872, + 0.0027865807060152292, + 0.003297991817817092, + 0.001495508593507111, + 0.0020226736087352037, + 0.001755838398821652, + 0.001418213709257543, + 0.0020820628851652145, + 0.0025746054016053677, + 0.0019416710129007697, + 0.002257330808788538, + 0.004410227760672569, + 0.004358806647360325, + 0.0050389207899570465, + 0.0048502967692911625, + 0.004064737819135189, + 0.006182743236422539, + 0.005167242605239153, + 0.0039042222779244184, + 0.005733927711844444, + 0.009506545960903168, + 0.006886763032525778, + 0.013205457478761673, + 0.007644787430763245, + 0.01083680335432291, + 0.0060195764526724815, + 0.012208198197185993, + 0.005778693128377199, + 0.011638693511486053, + 0.01764863356947899, + 0.008693806827068329, + 0.006161637604236603, + 0.007703454699367285, + 0.006221781950443983, + 0.00860036164522171, + 0.05591579154133797, + 0.047442030161619186, + 0.24084611237049103, + 0.0470786988735199, + 0.020980650559067726, + 0.10404396057128906, + 0.03438003733754158, + 0.025230687111616135, + 0.00264982832595706, + 0.0029446613043546677, + 0.003029311541467905, + 0.00485986890271306, + 0.0036051799543201923, + 0.002647766377776861, + 0.0026645888574421406, + 0.0028601926751434803, + 0.006925786379724741, + 0.007225405424833298, + 0.009914525784552097, + 0.005606206599622965, + 0.005464231129735708, + 0.0038754334673285484, + 0.0037558472249656916, + 0.004692287649959326, + 0.0028431478422135115, + 0.00318935071118176, + 0.0036512783262878656, + 0.0036244401708245277, + 0.004002892412245274, + 0.0021701673977077007, + 0.0034639588557183743, + 0.002584337256848812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.2015765756368637 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022004456259310246 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.0022004456259310246, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022004456259310246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14845019578933716, + 0.09176409989595413, + 0.058533623814582825, + 0.1070248931646347, + 0.09703879803419113, + 0.08258777856826782, + 0.20256732404232025, + 0.04462988302111626, + 0.07399404793977737, + 0.09340942651033401 + ], + "credit_map": [ + 0.01713581010699272, + 0.013644460588693619, + 0.02036452479660511, + 0.020992625504732132, + 0.014449696987867355, + 0.025482522323727608, + 0.02029372565448284, + 0.01608680561184883, + 0.009304151870310307, + 0.007325999904423952, + 0.015121733769774437, + 0.013446563854813576, + 0.009678083471953869, + 0.01661611907184124, + 0.011600791476666927, + 0.008670641109347343, + 0.0060622612945735455, + 0.008239061571657658, + 0.004786740988492966, + 0.006294239312410355, + 0.009926593862473965, + 0.009384111501276493, + 0.006892568897455931, + 0.0069480412639677525, + 0.011858785524964333, + 0.013062027283012867, + 0.011545626446604729, + 0.01721234805881977, + 0.014642247930169106, + 0.015923185274004936, + 0.010979198850691319, + 0.011801459826529026, + 0.009972730651497841, + 0.011021523736417294, + 0.010834894143044949, + 0.018683603033423424, + 0.01184755377471447, + 0.011840511113405228, + 0.010520092211663723, + 0.012317885644733906, + 0.006884532980620861, + 0.013189638964831829, + 0.009433969855308533, + 0.012221554294228554, + 0.008845875039696693, + 0.010919660329818726, + 0.010262023657560349, + 0.010830515995621681, + 0.0134546784684062, + 0.04170893132686615, + 0.05462929606437683, + 0.015065519139170647, + 0.011487663723528385, + 0.029917137697339058, + 0.014519773423671722, + 0.021784309297800064, + 0.00448306929320097, + 0.004791229963302612, + 0.005908075254410505, + 0.006783884484320879, + 0.008132684975862503, + 0.005233103409409523, + 0.0051946514286100864, + 0.004103176295757294, + 0.009907621890306473, + 0.009477301500737667, + 0.011763998307287693, + 0.01300289947539568, + 0.010629523545503616, + 0.005417025648057461, + 0.007738993037492037, + 0.006056679412722588, + 0.0106975631788373, + 0.01562380138784647, + 0.012711958959698677, + 0.012705959379673004, + 0.01178141962736845, + 0.006453329231590033, + 0.01271357573568821, + 0.010721806436777115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.0022004456259310246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809374198317528 + ], + "opposite_logprob_tokens": [ + -3.2880938053131104 + ], + "logprob_reward": -0.03809374198317528, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809374198317528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07712215185165405, + 0.035870641469955444, + 0.035202957689762115, + 0.02012208290398121, + 0.028333069756627083, + 0.07849740236997604, + 0.5633878707885742, + 0.06322091072797775, + 0.04402158781886101, + 0.05422130227088928 + ], + "credit_map": [ + 0.007728496100753546, + 0.0073865544982254505, + 0.011386370286345482, + 0.011192617937922478, + 0.007532807532697916, + 0.013412101194262505, + 0.010343743488192558, + 0.008139461278915405, + 0.003671657294034958, + 0.0037667909637093544, + 0.0043967105448246, + 0.003917772322893143, + 0.0034995016176253557, + 0.0066787307150661945, + 0.0037784709129482508, + 0.006161003839224577, + 0.002239784225821495, + 0.0032197677064687014, + 0.003892233595252037, + 0.0047354926355183125, + 0.004578553140163422, + 0.006404308136552572, + 0.005759629886597395, + 0.004373190458863974, + 0.0020081023685634136, + 0.0020906073041260242, + 0.0025374244432896376, + 0.003749767318367958, + 0.002370137022808194, + 0.0032600134145468473, + 0.0017502851551398635, + 0.002355745993554592, + 0.002531565958634019, + 0.004357919562608004, + 0.002672589384019375, + 0.004335824400186539, + 0.0032793302088975906, + 0.004602351225912571, + 0.002334184944629669, + 0.004219304770231247, + 0.005743337329477072, + 0.010767620988190174, + 0.019605448469519615, + 0.009598948992788792, + 0.006614820100367069, + 0.0070444513112306595, + 0.007719339802861214, + 0.011403435841202736, + 0.05330749601125717, + 0.05056887865066528, + 0.23151612281799316, + 0.044956739991903305, + 0.022642722353339195, + 0.10060854256153107, + 0.0352572463452816, + 0.024530166760087013, + 0.006287107244133949, + 0.010941616259515285, + 0.008139979094266891, + 0.009394130669534206, + 0.00668755266815424, + 0.005552203394472599, + 0.0054380628280341625, + 0.010780259035527706, + 0.005964518059045076, + 0.006549907848238945, + 0.00808756984770298, + 0.008078494109213352, + 0.00494874594733119, + 0.0029144021682441235, + 0.00374488509260118, + 0.003733070334419608, + 0.004967238288372755, + 0.00902277510613203, + 0.0068391370587050915, + 0.007376669906079769, + 0.007093791849911213, + 0.004231267608702183, + 0.007452309597283602, + 0.0072381142526865005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.03809374198317528 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide-angle drone shot, beginning at a low angle and slowly zooming out, reveals a section of the ancient Great Wall bathed in dramatic dusk light, perched atop a steep, rugged ridge. Constructed of weathered gray boulders and mottled blue bricks, the wall's surface is covered in age-worn cracks and moss. The setting sun casts long, twisting shadows across its rough surface. Suddenly, accompanied by a low tremor, a majestic beacon tower structure begins to shake violently, sending tiny pieces of gravel and dust cascading down from the cracks between the bricks. Immediately afterwards, an unstoppable chain reaction broke out. The huge brick and stone structure disintegrated in an instant. Tons of bricks and stones exploded and poured down like a landslide. The collapsed structure turned into a huge, rolling cloud of yellow-brown dust. This cloud of dust was dyed a fiery orange-red under the reflection of the setting sun and quickly engulfed the hillside below. The whole picture was filled with an epic sense of destruction and the ruthless power of time, creating a grand and tragic atmosphere.", + "seed": 27771, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_15/index_9549_seed_27771_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16909801899455487, + "vlm_logprob": -0.16909801899455487, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.16909801899455487, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5089309215545654 + ], + "opposite_logprob_tokens": [ + -1.0089309215545654 + ], + "logprob_reward": -0.5089309215545654, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.5089309215545654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12298686057329178, + 0.24641023576259613, + 0.03546140342950821, + 0.10402299463748932, + 0.0605730339884758, + 0.0584772489964962, + 0.13364365696907043, + 0.08649108558893204, + 0.05162196233868599, + 0.10031154751777649 + ], + "credit_map": [ + 0.008374536409974098, + 0.008627918548882008, + 0.02663281373679638, + 0.016973625868558884, + 0.0051338463090360165, + 0.007613126654177904, + 0.014760401099920273, + 0.034870583564043045, + 0.019420504570007324, + 0.008080917410552502, + 0.07765068113803864, + 0.04068738967180252, + 0.005828967783600092, + 0.01149821188300848, + 0.014398318715393543, + 0.0688452497124672, + 0.002420820062980056, + 0.0021174440626055002, + 0.009321577847003937, + 0.010987458750605583, + 0.0016425393987447023, + 0.001413423102349043, + 0.0047127013094723225, + 0.0028454402927309275, + 0.008270755410194397, + 0.004000531043857336, + 0.03213614970445633, + 0.015130020678043365, + 0.0026847089175134897, + 0.005036265589296818, + 0.006144666578620672, + 0.030619893223047256, + 0.004274238832294941, + 0.0032686174381524324, + 0.014310993254184723, + 0.006561247166246176, + 0.006201261188834906, + 0.0048667569644749165, + 0.008769082836806774, + 0.012320835143327713, + 0.0048004942946136, + 0.004481236916035414, + 0.012885971926152706, + 0.005376107059419155, + 0.004230258986353874, + 0.0052253915928304195, + 0.00666922889649868, + 0.014808561652898788, + 0.007923798635601997, + 0.010854935273528099, + 0.03849806264042854, + 0.042565274983644485, + 0.010563722811639309, + 0.0068869865499436855, + 0.01004236564040184, + 0.006308516953140497, + 0.01890621706843376, + 0.008687199093401432, + 0.008946023881435394, + 0.010764369741082191, + 0.008858388289809227, + 0.012401443906128407, + 0.010625302791595459, + 0.007302134297788143, + 0.010780886746942997, + 0.006010511424392462, + 0.0057466342113912106, + 0.008164890110492706, + 0.005006271880120039, + 0.005229954142123461, + 0.006942001171410084, + 0.003740807296708226, + 0.011874009855091572, + 0.018626296892762184, + 0.0096517289057374, + 0.015076307579874992, + 0.012136516161262989, + 0.012050032615661621, + 0.011628925800323486, + 0.009267736226320267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.5089309215545654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909954875707626 + ], + "opposite_logprob_tokens": [ + -2.464099645614624 + ], + "logprob_reward": -0.08909954875707626, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909954875707626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06743651628494263, + 0.08913546055555344, + 0.06037884205579758, + 0.10384383797645569, + 0.16896244883537292, + 0.2420281618833542, + 0.10207653790712357, + 0.03239010274410248, + 0.06458386033773422, + 0.0691642314195633 + ], + "credit_map": [ + 0.00815701112151146, + 0.006836161017417908, + 0.016030725091695786, + 0.012257968075573444, + 0.004878516308963299, + 0.003820279147475958, + 0.00661949859932065, + 0.008836359716951847, + 0.005534200929105282, + 0.003751849289983511, + 0.030498629435896873, + 0.024432159960269928, + 0.00204076012596488, + 0.0031048106029629707, + 0.00631710235029459, + 0.013455943204462528, + 0.003957039210945368, + 0.0019256547093391418, + 0.020767565816640854, + 0.013807586394250393, + 0.00129721499979496, + 0.002171195112168789, + 0.003988196142017841, + 0.012464389204978943, + 0.007976282387971878, + 0.00347234308719635, + 0.03359740972518921, + 0.01882440410554409, + 0.0023797452449798584, + 0.004419727250933647, + 0.006322839297354221, + 0.02685108780860901, + 0.011050096713006496, + 0.00882994756102562, + 0.041959136724472046, + 0.01810666173696518, + 0.012172391638159752, + 0.011649210937321186, + 0.025761881843209267, + 0.03943311423063278, + 0.018969763070344925, + 0.012722896412014961, + 0.06640482693910599, + 0.019574804231524467, + 0.010577712208032608, + 0.011584226973354816, + 0.02455977536737919, + 0.07763415575027466, + 0.005783971399068832, + 0.00602505961433053, + 0.030008623376488686, + 0.03861432150006294, + 0.005537013523280621, + 0.0033107309136539698, + 0.00646556681022048, + 0.006331247743219137, + 0.006931979209184647, + 0.003885403973981738, + 0.0033416852820664644, + 0.004686356522142887, + 0.003046449739485979, + 0.0034739987459033728, + 0.004466864746063948, + 0.0025573631282895803, + 0.013459023088216782, + 0.009961730800569057, + 0.008330252952873707, + 0.007773930672556162, + 0.0052495994605124, + 0.0057371086440980434, + 0.00869069155305624, + 0.00538152689114213, + 0.007331564091145992, + 0.00889662466943264, + 0.010540228337049484, + 0.009526428766548634, + 0.008603428490459919, + 0.008621054701507092, + 0.008396429941058159, + 0.007248477078974247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.08909954875707626 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015765756368637 + ], + "opposite_logprob_tokens": [ + -1.701576590538025 + ], + "logprob_reward": -0.2015765756368637, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015765756368637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061567485332489014, + 0.020632648840546608, + 0.16685710847377777, + 0.04395882040262222, + 0.049480680376291275, + 0.06385128200054169, + 0.4453393220901489, + 0.024408046156167984, + 0.05872654914855957, + 0.0651780366897583 + ], + "credit_map": [ + 0.008699340745806694, + 0.006304645445197821, + 0.009184897877275944, + 0.008104413747787476, + 0.005397084169089794, + 0.004439863841980696, + 0.006381438113749027, + 0.013055801391601562, + 0.0015176729066297412, + 0.0014657351421192288, + 0.004756744485348463, + 0.005784953013062477, + 0.0014506794977933168, + 0.0012646196410059929, + 0.001733423094265163, + 0.00265882327221334, + 0.010978512465953827, + 0.0045179626904428005, + 0.05870220437645912, + 0.03790377080440521, + 0.00307296309620142, + 0.006196568254381418, + 0.008955125696957111, + 0.036529991775751114, + 0.0035726153291761875, + 0.001999238505959511, + 0.012309282086789608, + 0.005618290975689888, + 0.0018486520275473595, + 0.002487852470949292, + 0.0029034134931862354, + 0.013219473883509636, + 0.004692467395216227, + 0.003954831976443529, + 0.006966156419366598, + 0.0067639728076756, + 0.007668151054531336, + 0.00737782334908843, + 0.006002611946314573, + 0.006054665893316269, + 0.004961593542248011, + 0.004028748255223036, + 0.01737712323665619, + 0.005164497997611761, + 0.0029052668251097202, + 0.0029594937805086374, + 0.006112762726843357, + 0.020341800525784492, + 0.023115407675504684, + 0.026916425675153732, + 0.11714313924312592, + 0.16075119376182556, + 0.03621651977300644, + 0.019850032404065132, + 0.026457207277417183, + 0.03488938882946968, + 0.003994036931544542, + 0.0025714843068271875, + 0.0026698389556258917, + 0.004244500305503607, + 0.001995476195588708, + 0.0028876697178930044, + 0.003553419141098857, + 0.0024916226975619793, + 0.00555737130343914, + 0.008072983473539352, + 0.006853682454675436, + 0.011865525506436825, + 0.005966439843177795, + 0.0056039015762507915, + 0.007788362912833691, + 0.00701828021556139, + 0.005214108154177666, + 0.00852628331631422, + 0.008127884939312935, + 0.013014706782996655, + 0.008491730317473412, + 0.004747869912534952, + 0.010120152495801449, + 0.006935299374163151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -0.2015765756368637 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820563269779086 + ], + "opposite_logprob_tokens": [ + -5.8778204917907715 + ], + "logprob_reward": -0.002820563269779086, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820563269779086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05255992338061333, + 0.033144354820251465, + 0.08917725831270218, + 0.0919409990310669, + 0.07553046196699142, + 0.09538906812667847, + 0.43519216775894165, + 0.03244508057832718, + 0.03384615108370781, + 0.06077451631426811 + ], + "credit_map": [ + 0.007545863278210163, + 0.005154245067387819, + 0.007160477805882692, + 0.007367582526057959, + 0.00612648343667388, + 0.005520553793758154, + 0.007144427392631769, + 0.00654028682038188, + 0.002188630634918809, + 0.0018665067618712783, + 0.010655066929757595, + 0.009036680683493614, + 0.0015893809031695127, + 0.0016826214268803596, + 0.0020147860050201416, + 0.00411068182438612, + 0.006107060704380274, + 0.0027737230993807316, + 0.030665617436170578, + 0.019982436671853065, + 0.002061113016679883, + 0.0034912065602838993, + 0.004949834197759628, + 0.01914627104997635, + 0.006773404777050018, + 0.0031303302384912968, + 0.028976576402783394, + 0.014020603150129318, + 0.0024360830429941416, + 0.004465848207473755, + 0.005723979789763689, + 0.026414168998599052, + 0.0063392831943929195, + 0.004528471268713474, + 0.014695721678435802, + 0.010975461453199387, + 0.007635049056261778, + 0.008120004087686539, + 0.012095739133656025, + 0.011140735819935799, + 0.007326903752982616, + 0.005477974656969309, + 0.024828724563121796, + 0.00747915031388402, + 0.005575224757194519, + 0.006271510850638151, + 0.009229405783116817, + 0.029200173914432526, + 0.025097990408539772, + 0.026577958837151527, + 0.13165521621704102, + 0.1645696759223938, + 0.024869564920663834, + 0.013053013011813164, + 0.0251776110380888, + 0.024191144853830338, + 0.006999457720667124, + 0.0035216105170547962, + 0.0035671480000019073, + 0.004652113653719425, + 0.003064539982005954, + 0.00307080359198153, + 0.004798242822289467, + 0.0027711649890989065, + 0.0041699944995343685, + 0.003774523502215743, + 0.005011490546166897, + 0.005622694734483957, + 0.004527227021753788, + 0.002812510123476386, + 0.005528817418962717, + 0.002398896962404251, + 0.005448898300528526, + 0.0076368399895727634, + 0.008374182507395744, + 0.009524669498205185, + 0.00919763557612896, + 0.004435621667653322, + 0.009588679298758507, + 0.006567990407347679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.002820563269779086 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430624857544899 + ], + "opposite_logprob_tokens": [ + -3.168062448501587 + ], + "logprob_reward": -0.0430624857544899, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430624857544899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11023341864347458, + 0.06938276439905167, + 0.18338879942893982, + 0.12260673195123672, + 0.07120343297719955, + 0.09714869409799576, + 0.1182841882109642, + 0.08826710283756256, + 0.047759268432855606, + 0.09172557294368744 + ], + "credit_map": [ + 0.014256788417696953, + 0.010459057986736298, + 0.022698301821947098, + 0.014423572458326817, + 0.007569900713860989, + 0.006748404819518328, + 0.010100536979734898, + 0.02397686243057251, + 0.0040499744936823845, + 0.003488735295832157, + 0.02364787645637989, + 0.022835973650217056, + 0.0023777505848556757, + 0.00270340614952147, + 0.00448384415358305, + 0.005795207340270281, + 0.012710429728031158, + 0.0060231066308915615, + 0.06251764297485352, + 0.038600143045186996, + 0.004001237917691469, + 0.0074027301743626595, + 0.010323996655642986, + 0.04180951043963432, + 0.009141769260168076, + 0.0043558343313634396, + 0.038738835602998734, + 0.019736068323254585, + 0.003858623094856739, + 0.005792077165096998, + 0.007548011839389801, + 0.033435504883527756, + 0.005072447005659342, + 0.003190127667039633, + 0.01824304461479187, + 0.007976661436259747, + 0.00504419906064868, + 0.005813723895698786, + 0.008914457634091377, + 0.016948768869042397, + 0.007306614425033331, + 0.006533580366522074, + 0.024851026013493538, + 0.008393430151045322, + 0.005543381441384554, + 0.005919085815548897, + 0.009455875493586063, + 0.02914569154381752, + 0.006415215320885181, + 0.008885275572538376, + 0.026698268949985504, + 0.03680107742547989, + 0.010964888148009777, + 0.007188003044575453, + 0.009186852723360062, + 0.012144614942371845, + 0.020100390538573265, + 0.009347167797386646, + 0.00951693020761013, + 0.011151126585900784, + 0.008218042552471161, + 0.009905434213578701, + 0.012067261151969433, + 0.007960746064782143, + 0.008709901012480259, + 0.004584303125739098, + 0.005834214389324188, + 0.00791570357978344, + 0.004881682805716991, + 0.0049939239397645, + 0.006903442088514566, + 0.003936095628887415, + 0.008397757075726986, + 0.01133855152875185, + 0.012063457630574703, + 0.0168208796530962, + 0.011604483239352703, + 0.007722897455096245, + 0.01271148957312107, + 0.011066052131354809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.0430624857544899 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide-angle drone shot, beginning at a low angle and slowly zooming out, reveals a section of the ancient Great Wall bathed in dramatic dusk light, perched atop a steep, rugged ridge. Constructed of weathered gray boulders and mottled blue bricks, the wall's surface is covered in age-worn cracks and moss. The setting sun casts long, twisting shadows across its rough surface. Suddenly, accompanied by a low tremor, a majestic beacon tower structure begins to shake violently, sending tiny pieces of gravel and dust cascading down from the cracks between the bricks. Immediately afterwards, an unstoppable chain reaction broke out. The huge brick and stone structure disintegrated in an instant. Tons of bricks and stones exploded and poured down like a landslide. The collapsed structure turned into a huge, rolling cloud of yellow-brown dust. This cloud of dust was dyed a fiery orange-red under the reflection of the setting sun and quickly engulfed the hillside below. The whole picture was filled with an epic sense of destruction and the ruthless power of time, creating a grand and tragic atmosphere.", + "seed": 27772, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_15/index_9549_seed_27772_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1927508189342916, + "vlm_logprob": -0.1927508189342916, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.1927508189342916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6031062602996826 + ], + "opposite_logprob_tokens": [ + -0.8531062602996826 + ], + "logprob_reward": -0.6031062602996826, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.6031062602996826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25513797998428345, + 0.10843503475189209, + 0.053917936980724335, + 0.16596658527851105, + 0.052488215267658234, + 0.058992985635995865, + 0.09014107286930084, + 0.06023220345377922, + 0.09289903938770294, + 0.06178898364305496 + ], + "credit_map": [ + 0.0338948555290699, + 0.04955759271979332, + 0.022799689322710037, + 0.01975741982460022, + 0.056704360991716385, + 0.035322751849889755, + 0.018667321652173996, + 0.01843397133052349, + 0.01292027160525322, + 0.016614742577075958, + 0.010496308095753193, + 0.008196196518838406, + 0.028864199295639992, + 0.015481198206543922, + 0.007776573766022921, + 0.008085531182587147, + 0.006337583996355534, + 0.007413262501358986, + 0.0038218002300709486, + 0.005715544801205397, + 0.011338654905557632, + 0.00744019215926528, + 0.007142739370465279, + 0.004708151798695326, + 0.023782003670930862, + 0.022813234478235245, + 0.01664712093770504, + 0.017488891258835793, + 0.02793153189122677, + 0.021405009552836418, + 0.015638567507267, + 0.020260199904441833, + 0.009208460338413715, + 0.009495463222265244, + 0.006028806325048208, + 0.006968783680349588, + 0.005970108788460493, + 0.0068329633213579655, + 0.004821399692445993, + 0.003162225242704153, + 0.008017455227673054, + 0.009507587179541588, + 0.006718240678310394, + 0.007092732936143875, + 0.0067829350009560585, + 0.007267496548593044, + 0.0062913293950259686, + 0.0073152026161551476, + 0.00888624507933855, + 0.015383381396532059, + 0.010473779402673244, + 0.021954117342829704, + 0.006822213064879179, + 0.012287582270801067, + 0.008482453413307667, + 0.005851289723068476, + 0.006575079169124365, + 0.004807972349226475, + 0.011946865357458591, + 0.006953765172511339, + 0.006375882774591446, + 0.00924895703792572, + 0.0074699013493955135, + 0.006853774189949036, + 0.013073812238872051, + 0.012780673801898956, + 0.018625322729349136, + 0.016573185101151466, + 0.007363514043390751, + 0.010446653701364994, + 0.007672401610761881, + 0.006363464053720236, + 0.007952256128191948, + 0.008662570267915726, + 0.01032545417547226, + 0.01133421529084444, + 0.007501589599996805, + 0.007698076777160168, + 0.004287000745534897, + 0.004027821123600006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.6031062602996826 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22564411163330078 + ], + "opposite_logprob_tokens": [ + -1.6006441116333008 + ], + "logprob_reward": -0.22564411163330078, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22564411163330078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24609243869781494, + 0.11658933013677597, + 0.06887495517730713, + 0.19053679704666138, + 0.07137864828109741, + 0.04050983116030693, + 0.0565667487680912, + 0.05865374580025673, + 0.06400640308856964, + 0.08679109811782837 + ], + "credit_map": [ + 0.04187827557325363, + 0.027745909988880157, + 0.031877461820840836, + 0.03003779798746109, + 0.03437088429927826, + 0.03339868411421776, + 0.024610159918665886, + 0.022173255681991577, + 0.015622500330209732, + 0.014273644424974918, + 0.015247244387865067, + 0.015716837719082832, + 0.015216854400932789, + 0.015217861160635948, + 0.01320025883615017, + 0.012094127014279366, + 0.009997370652854443, + 0.00792897678911686, + 0.007978987880051136, + 0.00705179525539279, + 0.012501556426286697, + 0.009145598858594894, + 0.006699774414300919, + 0.007570893503725529, + 0.02819804660975933, + 0.02447805181145668, + 0.016045456752181053, + 0.02169867232441902, + 0.031553272157907486, + 0.023057181388139725, + 0.024708405137062073, + 0.020797714591026306, + 0.009673449210822582, + 0.0066841463558375835, + 0.012620980851352215, + 0.00838969461619854, + 0.005779176950454712, + 0.010526616126298904, + 0.009517857804894447, + 0.00818672776222229, + 0.0047436063177883625, + 0.005027904640883207, + 0.004599364940077066, + 0.005963997449725866, + 0.005710883066058159, + 0.0052916607819497585, + 0.00478680245578289, + 0.004385610111057758, + 0.004831986967474222, + 0.009737465530633926, + 0.0053188009187579155, + 0.006600207183510065, + 0.005373557098209858, + 0.00725599005818367, + 0.00943005084991455, + 0.008018692024052143, + 0.009153394959867, + 0.009448950178921223, + 0.006999067962169647, + 0.00784224271774292, + 0.0058376542292535305, + 0.006811464671045542, + 0.006760342977941036, + 0.005800624378025532, + 0.007593333255499601, + 0.006758510135114193, + 0.011573434807360172, + 0.009129565209150314, + 0.007194971200078726, + 0.008114917203783989, + 0.007060220465064049, + 0.006581451743841171, + 0.008397704921662807, + 0.008841827511787415, + 0.015009746886789799, + 0.014875420369207859, + 0.008184650912880898, + 0.013863095082342625, + 0.00931823905557394, + 0.00830041617155075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.22564411163330078 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901136577129364 + ], + "opposite_logprob_tokens": [ + -2.5790114402770996 + ], + "logprob_reward": -0.07901136577129364, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901136577129364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09143733233213425, + 0.05191854387521744, + 0.02959417551755905, + 0.06074494123458862, + 0.09493861347436905, + 0.10912640392780304, + 0.16827741265296936, + 0.14943496882915497, + 0.17619723081588745, + 0.068330317735672 + ], + "credit_map": [ + 0.012547260150313377, + 0.013994760811328888, + 0.011647558771073818, + 0.009293108247220516, + 0.015384376049041748, + 0.013958192430436611, + 0.00905519351363182, + 0.005556885618716478, + 0.007811321411281824, + 0.008848784491419792, + 0.005942509975284338, + 0.004901289474219084, + 0.008365082554519176, + 0.006780365481972694, + 0.004706579726189375, + 0.004562611225992441, + 0.004155599046498537, + 0.004796518012881279, + 0.002769834827631712, + 0.0020598028786480427, + 0.006960200611501932, + 0.00416101748123765, + 0.002376190386712551, + 0.0023150129709392786, + 0.006197001785039902, + 0.010070678777992725, + 0.005283563397824764, + 0.006228675600141287, + 0.01241173967719078, + 0.008315155282616615, + 0.007394540589302778, + 0.004843584727495909, + 0.012830066494643688, + 0.01057203859090805, + 0.0156423207372427, + 0.011489522643387318, + 0.010986664332449436, + 0.013608395121991634, + 0.010949065908789635, + 0.008860534057021141, + 0.020455433055758476, + 0.02101285569369793, + 0.012743985280394554, + 0.010904517956078053, + 0.012755300849676132, + 0.01310182735323906, + 0.009825153276324272, + 0.008327336050570011, + 0.019221071153879166, + 0.024423064664006233, + 0.01902482844889164, + 0.03965591639280319, + 0.013926757499575615, + 0.021812764927744865, + 0.018110061064362526, + 0.012102958746254444, + 0.02122003771364689, + 0.026315508410334587, + 0.016862796619534492, + 0.022596774622797966, + 0.016201263293623924, + 0.017262771725654602, + 0.01827346347272396, + 0.010702351108193398, + 0.02322026900947094, + 0.02306751161813736, + 0.03250901773571968, + 0.032314881682395935, + 0.01724645309150219, + 0.02281929738819599, + 0.012294243089854717, + 0.012725558131933212, + 0.008203973062336445, + 0.007154403254389763, + 0.011795960366725922, + 0.010404138825833797, + 0.008223924785852432, + 0.009220798499882221, + 0.007323811296373606, + 0.006003313232213259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.07901136577129364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010432521812617779 + ], + "opposite_logprob_tokens": [ + -6.876043319702148 + ], + "logprob_reward": -0.0010432521812617779, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010432521812617779, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13207805156707764, + 0.06431756913661957, + 0.05915489047765732, + 0.060959137976169586, + 0.11747898906469345, + 0.07716093957424164, + 0.11995916813611984, + 0.11019797623157501, + 0.08922911435365677, + 0.1694640964269638 + ], + "credit_map": [ + 0.011218657717108727, + 0.02632683515548706, + 0.013463367708027363, + 0.011413107626140118, + 0.029671434313058853, + 0.01833394169807434, + 0.01178547739982605, + 0.0098652308806777, + 0.008275728672742844, + 0.011640307493507862, + 0.00457135122269392, + 0.004606733098626137, + 0.01764461025595665, + 0.00948351714760065, + 0.003704668488353491, + 0.004390649031847715, + 0.008487368002533913, + 0.008660029619932175, + 0.004838865250349045, + 0.005840417928993702, + 0.012896244414150715, + 0.008130309171974659, + 0.005613465793430805, + 0.004688190296292305, + 0.007990602403879166, + 0.008733530528843403, + 0.007309888955205679, + 0.00650435034185648, + 0.009417097084224224, + 0.008103924803435802, + 0.005630736704915762, + 0.00726900901645422, + 0.018119826912879944, + 0.017713187262415886, + 0.017788242548704147, + 0.011684400960803032, + 0.015190182253718376, + 0.016290443018078804, + 0.011034886352717876, + 0.009657824411988258, + 0.014044392853975296, + 0.013778621330857277, + 0.00866005290299654, + 0.007983401417732239, + 0.008521940559148788, + 0.009456883184611797, + 0.008566539734601974, + 0.006149106193333864, + 0.01171954721212387, + 0.016019143164157867, + 0.012526651844382286, + 0.029011709615588188, + 0.010524716228246689, + 0.01624930091202259, + 0.014166138134896755, + 0.009741967543959618, + 0.015076358802616596, + 0.013605612330138683, + 0.01511138491332531, + 0.014436199329793453, + 0.012075971812009811, + 0.01431281492114067, + 0.015566132962703705, + 0.010013503022491932, + 0.010732503607869148, + 0.01343573722988367, + 0.01212400570511818, + 0.015602258034050465, + 0.00853294599801302, + 0.010401192121207714, + 0.010716114193201065, + 0.007684353273361921, + 0.027094466611742973, + 0.025312349200248718, + 0.030992252752184868, + 0.021899422630667686, + 0.02076011151075363, + 0.01795344054698944, + 0.01210881769657135, + 0.01334321964532137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0010432521812617779 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494910478591919 + ], + "opposite_logprob_tokens": [ + -2.9299490451812744 + ], + "logprob_reward": -0.05494910478591919, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494910478591919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2096092253923416, + 0.10892380028963089, + 0.07281631976366043, + 0.046466343104839325, + 0.04765710234642029, + 0.07375284284353256, + 0.08660272508859634, + 0.07169123739004135, + 0.1428578943014145, + 0.13962243497371674 + ], + "credit_map": [ + 0.021149272099137306, + 0.03682533651590347, + 0.019546039402484894, + 0.02378764934837818, + 0.04288602992892265, + 0.027731359004974365, + 0.01922859065234661, + 0.018454961478710175, + 0.012280677445232868, + 0.017071068286895752, + 0.011723130941390991, + 0.011712761595845222, + 0.02126001939177513, + 0.014906065538525581, + 0.010420258156955242, + 0.00954983476549387, + 0.008216893300414085, + 0.007977455854415894, + 0.012012208811938763, + 0.01030756626278162, + 0.009794588200747967, + 0.009445583447813988, + 0.008599134162068367, + 0.006462900433689356, + 0.004491034895181656, + 0.008191554807126522, + 0.004382866434752941, + 0.004271745681762695, + 0.009411615319550037, + 0.007076607551425695, + 0.0049948422238230705, + 0.0036460785195231438, + 0.005374089814722538, + 0.004658023826777935, + 0.007800501771271229, + 0.006755510810762644, + 0.005769165698438883, + 0.006579837761819363, + 0.006261653266847134, + 0.0044583226554095745, + 0.013626368716359138, + 0.013711446896195412, + 0.007435626350343227, + 0.007031452376395464, + 0.007099424954503775, + 0.009291796945035458, + 0.008519071154296398, + 0.007037656381726265, + 0.008742225356400013, + 0.01267187763005495, + 0.013726248405873775, + 0.016695881262421608, + 0.007968739606440067, + 0.011790723539888859, + 0.009112166240811348, + 0.005894870031625032, + 0.007075750734657049, + 0.007779395207762718, + 0.009144587442278862, + 0.008483347482979298, + 0.010682146064937115, + 0.010175204835832119, + 0.010594801045954227, + 0.0077560171484947205, + 0.02062731422483921, + 0.018944744020700455, + 0.03273593634366989, + 0.025900818407535553, + 0.010365321300923824, + 0.01587063819169998, + 0.009313281625509262, + 0.009099851362407207, + 0.01942308247089386, + 0.018052827566862106, + 0.028152091428637505, + 0.021495338529348373, + 0.014604056254029274, + 0.016808083280920982, + 0.010352256707847118, + 0.010734716430306435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.05494910478591919 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic wide-angle drone shot, beginning at a low angle and slowly zooming out, reveals a section of the ancient Great Wall bathed in dramatic dusk light, perched atop a steep, rugged ridge. Constructed of weathered gray boulders and mottled blue bricks, the wall's surface is covered in age-worn cracks and moss. The setting sun casts long, twisting shadows across its rough surface. Suddenly, accompanied by a low tremor, a majestic beacon tower structure begins to shake violently, sending tiny pieces of gravel and dust cascading down from the cracks between the bricks. Immediately afterwards, an unstoppable chain reaction broke out. The huge brick and stone structure disintegrated in an instant. Tons of bricks and stones exploded and poured down like a landslide. The collapsed structure turned into a huge, rolling cloud of yellow-brown dust. This cloud of dust was dyed a fiery orange-red under the reflection of the setting sun and quickly engulfed the hillside below. The whole picture was filled with an epic sense of destruction and the ruthless power of time, creating a grand and tragic atmosphere.", + "seed": 27773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_15/index_9549_seed_27773_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09380002075340599, + "vlm_logprob": -0.09380002075340599, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.09380002075340599, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34622740745544434 + ], + "opposite_logprob_tokens": [ + -1.3462274074554443 + ], + "logprob_reward": -0.34622740745544434, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.34622740745544434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23283028602600098, + 0.08006405085325241, + 0.08814818412065506, + 0.10423942655324936, + 0.0657479390501976, + 0.08216793090105057, + 0.24459396302700043, + 0.030042532831430435, + 0.03355073556303978, + 0.038614921271800995 + ], + "credit_map": [ + 0.011265858076512814, + 0.01217013318091631, + 0.017645426094532013, + 0.02592969685792923, + 0.01122364029288292, + 0.027427958324551582, + 0.07495653629302979, + 0.052211061120033264, + 0.002581005683168769, + 0.002271624282002449, + 0.004530842415988445, + 0.003798064077273011, + 0.0045934077352285385, + 0.005369093269109726, + 0.030836110934615135, + 0.02608390711247921, + 0.004070337396115065, + 0.0052292197942733765, + 0.005850963294506073, + 0.009560315869748592, + 0.010136420838534832, + 0.010038256645202637, + 0.02468499168753624, + 0.018577681854367256, + 0.00797580648213625, + 0.008288667537271976, + 0.0151592418551445, + 0.01406314130872488, + 0.007587301544845104, + 0.009506508708000183, + 0.015070701949298382, + 0.026588061824440956, + 0.00589728495106101, + 0.009717206470668316, + 0.009256009943783283, + 0.010870271362364292, + 0.005649339873343706, + 0.011788347736001015, + 0.005539748817682266, + 0.007029735017567873, + 0.007361935451626778, + 0.00591384619474411, + 0.013521982356905937, + 0.014615732245147228, + 0.005593861918896437, + 0.0125929806381464, + 0.012004101648926735, + 0.010563495568931103, + 0.02493450604379177, + 0.02141261287033558, + 0.051467061042785645, + 0.03416529670357704, + 0.026334552094340324, + 0.033926207572221756, + 0.021646643057465553, + 0.030707083642482758, + 0.0019549718126654625, + 0.003783876309171319, + 0.0032001477666199207, + 0.003948704805225134, + 0.002418670803308487, + 0.005491056479513645, + 0.005895327311009169, + 0.003349780570715666, + 0.0021937056444585323, + 0.0014364008093252778, + 0.003311177948489785, + 0.002318170154467225, + 0.0014695234131067991, + 0.008694866672158241, + 0.0036438354291021824, + 0.01048305630683899, + 0.002837828826159239, + 0.0020844440441578627, + 0.007144067902117968, + 0.00401994027197361, + 0.002571052173152566, + 0.006844113580882549, + 0.004654707387089729, + 0.008458771742880344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.34622740745544434 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048701103776693344 + ], + "opposite_logprob_tokens": [ + -3.048701047897339 + ], + "logprob_reward": -0.048701103776693344, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048701103776693344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16658972203731537, + 0.17111289501190186, + 0.07788629084825516, + 0.11112301796674728, + 0.06020567938685417, + 0.11143362522125244, + 0.06759560108184814, + 0.09362001717090607, + 0.0800347626209259, + 0.0603983998298645 + ], + "credit_map": [ + 0.01042720302939415, + 0.014989940449595451, + 0.024387454614043236, + 0.02782798372209072, + 0.024886419996619225, + 0.023115642368793488, + 0.019168751314282417, + 0.02178633213043213, + 0.007822527550160885, + 0.007856491021811962, + 0.012658180668950081, + 0.014844309538602829, + 0.01069845911115408, + 0.015782736241817474, + 0.0516371950507164, + 0.04981299489736557, + 0.0051113213412463665, + 0.005988082382827997, + 0.0099214231595397, + 0.013971937820315361, + 0.006592174991965294, + 0.014700456522405148, + 0.010544993914663792, + 0.011055904440581799, + 0.006012182217091322, + 0.00927689764648676, + 0.010954214259982109, + 0.01089547760784626, + 0.008518178015947342, + 0.011827369220554829, + 0.020632248371839523, + 0.0330064594745636, + 0.005768269766122103, + 0.008016899228096008, + 0.0072394926100969315, + 0.009650521911680698, + 0.0062805661000311375, + 0.009228136390447617, + 0.005278237629681826, + 0.00874356273561716, + 0.008147499524056911, + 0.008418631739914417, + 0.022346923127770424, + 0.018801316618919373, + 0.006382372695952654, + 0.01848035492002964, + 0.017837343737483025, + 0.011019181460142136, + 0.0062922523356974125, + 0.006307349074631929, + 0.01233329251408577, + 0.013156267814338207, + 0.006450697313994169, + 0.008894911035895348, + 0.006716971285641193, + 0.007443862035870552, + 0.005261008627712727, + 0.008171230554580688, + 0.014379242435097694, + 0.009639625437557697, + 0.007271869573742151, + 0.017375759780406952, + 0.018520597368478775, + 0.01300068385899067, + 0.005935710854828358, + 0.0034532586578279734, + 0.010512596927583218, + 0.008282518945634365, + 0.004736126866191626, + 0.012036969885230064, + 0.017636483535170555, + 0.017441103234887123, + 0.004978178068995476, + 0.00301431049592793, + 0.009762703441083431, + 0.008674903772771358, + 0.0043457574211061, + 0.009450365789234638, + 0.007313055451959372, + 0.012859121896326542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.048701103776693344 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054968852549791336 + ], + "opposite_logprob_tokens": [ + -2.92996883392334 + ], + "logprob_reward": -0.054968852549791336, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054968852549791336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08709599822759628, + 0.15217237174510956, + 0.03958066925406456, + 0.14825837314128876, + 0.1098397895693779, + 0.24501778185367584, + 0.0659673884510994, + 0.032999616116285324, + 0.045987095683813095, + 0.0730808675289154 + ], + "credit_map": [ + 0.005376330576837063, + 0.006785260979086161, + 0.010672089643776417, + 0.009136752225458622, + 0.015912014991044998, + 0.01126722153276205, + 0.01683693379163742, + 0.01110940333455801, + 0.0055260248482227325, + 0.006225940305739641, + 0.008919734507799149, + 0.006884031929075718, + 0.012647309340536594, + 0.013071754947304726, + 0.05531490966677666, + 0.04358268156647682, + 0.0018654243322089314, + 0.002488309284672141, + 0.0031376490369439125, + 0.003357598325237632, + 0.004553759004920721, + 0.0036932246293872595, + 0.012249833904206753, + 0.008234875276684761, + 0.009228053502738476, + 0.014849980361759663, + 0.01697739213705063, + 0.014188426546752453, + 0.016234703361988068, + 0.015380728989839554, + 0.022553350776433945, + 0.038845743983983994, + 0.010041932575404644, + 0.01326922606676817, + 0.01485355757176876, + 0.013286345638334751, + 0.014638610184192657, + 0.016345640644431114, + 0.013027478940784931, + 0.014377012848854065, + 0.012183619663119316, + 0.017516564577817917, + 0.06466532498598099, + 0.03900374472141266, + 0.013692634180188179, + 0.03193838149309158, + 0.04437592253088951, + 0.02164161205291748, + 0.004704032093286514, + 0.005034144502133131, + 0.013375567272305489, + 0.007440030574798584, + 0.005233106203377247, + 0.010490036569535732, + 0.005930958315730095, + 0.013759516179561615, + 0.0021975652780383825, + 0.0026565196458250284, + 0.005852636881172657, + 0.005381540395319462, + 0.0024455534294247627, + 0.005644695833325386, + 0.004759788978844881, + 0.0040613203309476376, + 0.0030473112128674984, + 0.0025272953789681196, + 0.007062507793307304, + 0.006677179131656885, + 0.002355805365368724, + 0.007672299165278673, + 0.007092077285051346, + 0.009552625939249992, + 0.004326359368860722, + 0.003259825287386775, + 0.012646708637475967, + 0.00820143148303032, + 0.004642503336071968, + 0.011721915565431118, + 0.01184634119272232, + 0.016435790807008743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.054968852549791336 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009205871028825641 + ], + "opposite_logprob_tokens": [ + -7.00092077255249 + ], + "logprob_reward": -0.0009205871028825641, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009205871028825641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15627513825893402, + 0.14969833195209503, + 0.18116503953933716, + 0.05998442322015762, + 0.10264185070991516, + 0.07448702305555344, + 0.10427077859640121, + 0.04780280217528343, + 0.06414112448692322, + 0.05953349173069 + ], + "credit_map": [ + 0.007904541678726673, + 0.009199016727507114, + 0.012772980146110058, + 0.019423626363277435, + 0.010153649374842644, + 0.0221117977052927, + 0.043879538774490356, + 0.030829979106783867, + 0.004288743250072002, + 0.0036809309385716915, + 0.007352301385253668, + 0.00864715501666069, + 0.006974703632295132, + 0.009454997256398201, + 0.05801771953701973, + 0.05128178000450134, + 0.005643061827868223, + 0.0039431508630514145, + 0.00892772525548935, + 0.009156069718301296, + 0.008039002306759357, + 0.011892749927937984, + 0.07232984155416489, + 0.061233438551425934, + 0.005070493556559086, + 0.006380893290042877, + 0.009751624427735806, + 0.007046393118798733, + 0.0051751211285591125, + 0.005356042645871639, + 0.009361803531646729, + 0.011842053383588791, + 0.009127041324973106, + 0.012459424324333668, + 0.012405846267938614, + 0.02068614400923252, + 0.00818855594843626, + 0.013868319801986217, + 0.009979761205613613, + 0.015926748514175415, + 0.005795351229608059, + 0.008611718192696571, + 0.010688519105315208, + 0.01358221285045147, + 0.006134456023573875, + 0.014186043292284012, + 0.007281880360096693, + 0.008206836879253387, + 0.009530077688395977, + 0.0071322922594845295, + 0.02294931374490261, + 0.013241633772850037, + 0.009414023719727993, + 0.01785646751523018, + 0.007979189977049828, + 0.0161677785217762, + 0.003271859372034669, + 0.0033712550066411495, + 0.006962715648114681, + 0.005394714884459972, + 0.0028474668506532907, + 0.012167656794190407, + 0.004701309837400913, + 0.009085825644433498, + 0.005054572597146034, + 0.003845152212306857, + 0.008277342654764652, + 0.008173417299985886, + 0.004400292411446571, + 0.0075532663613557816, + 0.014477679505944252, + 0.012359404936432838, + 0.003550478955730796, + 0.002552680904045701, + 0.011135287582874298, + 0.008090156130492687, + 0.003268497996032238, + 0.007400502450764179, + 0.010973325930535793, + 0.01256256178021431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0009205871028825641 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a wide-angle drone shot that begins at a low angle and slowly zooms out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Great Wall constructed of weathered gray boulders and mottled blue bricks with visible age-worn cracks and moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a beacon tower structure begin to shake violently, sending gravel and dust cascading down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the structure disintegrate instantly into a huge, rolling cloud of yellow-brown dust that turns fiery orange-red under the setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the dust cloud quickly engulf the hillside below?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818215288221836 + ], + "opposite_logprob_tokens": [ + -4.018182277679443 + ], + "logprob_reward": -0.01818215288221836, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818215288221836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19499610364437103, + 0.1281910538673401, + 0.10456951707601547, + 0.032849423587322235, + 0.07466939836740494, + 0.06882722675800323, + 0.08428000658750534, + 0.06642522662878036, + 0.19713975489139557, + 0.04805223271250725 + ], + "credit_map": [ + 0.008608429692685604, + 0.006831729784607887, + 0.012157770805060863, + 0.010692840442061424, + 0.011935810558497906, + 0.01321296114474535, + 0.07604973763227463, + 0.05550685152411461, + 0.004028893541544676, + 0.0035908769350498915, + 0.006351014133542776, + 0.00529448501765728, + 0.007416353560984135, + 0.008105160668492317, + 0.05001719668507576, + 0.04338708147406578, + 0.003339421469718218, + 0.0025545901153236628, + 0.005411537364125252, + 0.005690808407962322, + 0.004990648478269577, + 0.006622740533202887, + 0.041297923773527145, + 0.03466185927391052, + 0.0028493020217865705, + 0.0036580173764377832, + 0.005040939897298813, + 0.004018151666969061, + 0.0027240319177508354, + 0.002973968628793955, + 0.005194232799112797, + 0.006390783004462719, + 0.006385036278516054, + 0.005903533659875393, + 0.010775688104331493, + 0.013295448385179043, + 0.005472027231007814, + 0.011514217592775822, + 0.006196874193847179, + 0.01512658316642046, + 0.0048474823124706745, + 0.007770756259560585, + 0.007240637671202421, + 0.011231651529669762, + 0.005337797105312347, + 0.01338223833590746, + 0.011911620385944843, + 0.007105045951902866, + 0.007358557544648647, + 0.006329155992716551, + 0.017925407737493515, + 0.011050152592360973, + 0.007585069630295038, + 0.01311239879578352, + 0.006984094623476267, + 0.013935177586972713, + 0.004660711158066988, + 0.0053953323513269424, + 0.009892718866467476, + 0.009008437395095825, + 0.004640968982130289, + 0.008146308362483978, + 0.013875680044293404, + 0.010805071331560612, + 0.01068728044629097, + 0.007947300560772419, + 0.0167540293186903, + 0.013467705808579922, + 0.008182107470929623, + 0.05528385564684868, + 0.020901450887322426, + 0.06391603499650955, + 0.0033861559350043535, + 0.0021384628489613533, + 0.006909215822815895, + 0.004235607106238604, + 0.002955181524157524, + 0.009159908629953861, + 0.006152274087071419, + 0.013115430250763893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.01818215288221836 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_2/index_tensor([6534, 6534])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_2/index_tensor([6534, 6534])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9f8dd735c46ce47aa6c6bd10238dc2fd4c010af0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_2/index_tensor([6534, 6534])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly rotating clockwise medium shot, filmed from a slightly upward perspective, captures a mythical phoenix. Each of its feathers appears to be forged from pure liquid gold, radiating a warm, dazzling glow. This light flows slowly through its wings like molten lava, while occasionally tiny golden sparks flake from the wings' edges, scattering like stardust against the deep, dark backdrop of cosmic nebulae. The phoenix is in the climax of a graceful movement, fully extending its massive, shimmering wings, nearly filling the entire frame. Its slender tail traces a brilliant arc of golden light behind it. Its body executes a balletic rotation in a steady and leisurely rhythm, its majestic, slender crown slightly raised, its blazing ruby eyes gazing into the distance. The entire scene is enveloped in an atmosphere of sacredness, tranquility, and awe. The dominant colors are brilliant gold and deep black, punctuated by faint, distant shimmers of purple and blue starlight in the background, evoking a mythical creation scene from the dawn of the universe.", + "seed": 8715, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_2/index_6534_seed_8715_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.028128771856427193, + "vlm_logprob": -0.028128771856427193, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.028128771856427193, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0329911969602108 + ], + "opposite_logprob_tokens": [ + -3.5329911708831787 + ], + "logprob_reward": -0.0329911969602108, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0329911969602108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.663109540939331, + 0.081484355032444, + 0.018222810700535774, + 0.015022153034806252, + 0.049719542264938354, + 0.0271317046135664, + 0.022059893235564232, + 0.02398073859512806, + 0.03947637230157852, + 0.059792861342430115 + ], + "credit_map": [ + 0.24227862060070038, + 0.03452225774526596, + 0.04380152374505997, + 0.016260989010334015, + 0.027810536324977875, + 0.20546872913837433, + 0.06552102416753769, + 0.027445852756500244, + 0.0053849671967327595, + 0.004129142500460148, + 0.00461691664531827, + 0.0036422333214432, + 0.005723914597183466, + 0.02429003082215786, + 0.01113169826567173, + 0.022565459832549095, + 0.0035876475740224123, + 0.0015065876068547368, + 0.0019465904915705323, + 0.002312627388164401, + 0.001996053848415613, + 0.002732884371653199, + 0.0012672195443883538, + 0.002873200224712491, + 0.0016884008655324578, + 0.0019263917347416282, + 0.002220391295850277, + 0.0027218086179345846, + 0.0013959748903289437, + 0.0012871286598965526, + 0.0016457741148769855, + 0.002136283554136753, + 0.015070604160428047, + 0.00683841947466135, + 0.004435935989022255, + 0.004109813831746578, + 0.00534960487857461, + 0.0026917941868305206, + 0.0036178831942379475, + 0.007605484686791897, + 0.005638393573462963, + 0.0027254337910562754, + 0.002751278458163142, + 0.00365433213301003, + 0.0033579575829207897, + 0.0018497828859835863, + 0.003112372010946274, + 0.004042155109345913, + 0.0027899155393242836, + 0.0021094889380037785, + 0.002359488746151328, + 0.002267667558044195, + 0.0025503046344965696, + 0.002369350055232644, + 0.003652659012004733, + 0.0039610182866454124, + 0.003986741416156292, + 0.002406161744147539, + 0.0023944114800542593, + 0.0024983661714941263, + 0.002701075281947851, + 0.0038075814954936504, + 0.0021140819881111383, + 0.004072321113198996, + 0.009204388596117496, + 0.004620466381311417, + 0.00321777886711061, + 0.0027306920383125544, + 0.003218045923858881, + 0.0063746050000190735, + 0.0021980670280754566, + 0.007912327535450459, + 0.01126133929938078, + 0.005234875716269016, + 0.009079736657440662, + 0.007623742800205946, + 0.0056654284708201885, + 0.0066625443287193775, + 0.005983777344226837, + 0.008281414397060871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0329911969602108 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011096905916929245 + ], + "opposite_logprob_tokens": [ + -4.511096954345703 + ], + "logprob_reward": -0.011096905916929245, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011096905916929245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43373918533325195, + 0.13146281242370605, + 0.024967407807707787, + 0.07865595072507858, + 0.03620774671435356, + 0.05600573122501373, + 0.07244593650102615, + 0.03623078390955925, + 0.06904257833957672, + 0.061241887509822845 + ], + "credit_map": [ + 0.15450254082679749, + 0.023299939930438995, + 0.026999257504940033, + 0.0163872092962265, + 0.02215305157005787, + 0.1259019374847412, + 0.04452325776219368, + 0.019972000271081924, + 0.007775698788464069, + 0.007578661199659109, + 0.008101826533675194, + 0.0055060964077711105, + 0.009821989573538303, + 0.03906436264514923, + 0.01837669126689434, + 0.03523748368024826, + 0.003984733484685421, + 0.0028594392351806164, + 0.0024786926805973053, + 0.003522118553519249, + 0.003142566652968526, + 0.002559822518378496, + 0.002598484745249152, + 0.003821547608822584, + 0.009444601833820343, + 0.008863644674420357, + 0.00980968028306961, + 0.010469717904925346, + 0.009209728799760342, + 0.008413654752075672, + 0.01145116426050663, + 0.010993754491209984, + 0.0048334174789488316, + 0.0053757862187922, + 0.0038351300172507763, + 0.005277720745652914, + 0.004360495135188103, + 0.0030079372227191925, + 0.00438874214887619, + 0.005128515884280205, + 0.007593618705868721, + 0.004902665503323078, + 0.006038540508598089, + 0.008554602973163128, + 0.01034485548734665, + 0.00469278683885932, + 0.006656366400420666, + 0.00722229341045022, + 0.010430173017084599, + 0.005802045576274395, + 0.005938128102570772, + 0.007932170294225216, + 0.006871738005429506, + 0.009002001956105232, + 0.013277318328619003, + 0.01319235097616911, + 0.0038599215913563967, + 0.0030933658126741648, + 0.004163165111094713, + 0.005061398725956678, + 0.004166432190686464, + 0.006856979336589575, + 0.004779563285410404, + 0.004249953664839268, + 0.01363955158740282, + 0.006400430575013161, + 0.005877654999494553, + 0.004682756029069424, + 0.005581273697316647, + 0.0121083352714777, + 0.005885551683604717, + 0.014867017976939678, + 0.005473364610224962, + 0.005519666243344545, + 0.009510915726423264, + 0.00842815451323986, + 0.008133959025144577, + 0.008043495938181877, + 0.0068514300510287285, + 0.00928089302033186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.011096905916929245 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370063751935959 + ], + "opposite_logprob_tokens": [ + -3.408700704574585 + ], + "logprob_reward": -0.03370063751935959, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370063751935959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1803998202085495, + 0.048053983598947525, + 0.03284282982349396, + 0.043641891330480576, + 0.08326277136802673, + 0.033029429614543915, + 0.04489004984498024, + 0.059932149946689606, + 0.4220938980579376, + 0.05185314640402794 + ], + "credit_map": [ + 0.02198699116706848, + 0.025479594245553017, + 0.01751169189810753, + 0.026527579873800278, + 0.018717488273978233, + 0.016941314563155174, + 0.031215326860547066, + 0.022019827738404274, + 0.005104103125631809, + 0.004259321838617325, + 0.0071911015547811985, + 0.004502380732446909, + 0.004728158935904503, + 0.006478355266153812, + 0.007186539936810732, + 0.0086040198802948, + 0.003926438745111227, + 0.003866377053782344, + 0.0030957190319895744, + 0.002999258926138282, + 0.004048606380820274, + 0.004117227625101805, + 0.0051746489480137825, + 0.0056145526468753815, + 0.003248984692618251, + 0.004704758059233427, + 0.004808981437236071, + 0.006564275361597538, + 0.004931767471134663, + 0.0036832536570727825, + 0.008305530063807964, + 0.007394342217594385, + 0.017081409692764282, + 0.011616184376180172, + 0.008893576450645924, + 0.011709135957062244, + 0.011185661889612675, + 0.005636101588606834, + 0.006605864968150854, + 0.010534830391407013, + 0.0030191296245902777, + 0.0029230983927845955, + 0.004310807213187218, + 0.005774843972176313, + 0.005209941416978836, + 0.003913139924407005, + 0.0033803321421146393, + 0.004498135764151812, + 0.004717678762972355, + 0.00479746051132679, + 0.003844099584966898, + 0.0044766319915652275, + 0.005990972276777029, + 0.00461072800680995, + 0.008028734475374222, + 0.008423744700849056, + 0.006156240124255419, + 0.004717202857136726, + 0.0056105018593370914, + 0.007831936702132225, + 0.00891665555536747, + 0.010704045183956623, + 0.006461688317358494, + 0.009533879347145557, + 0.04150579497218132, + 0.0300679299980402, + 0.04393854737281799, + 0.04392337426543236, + 0.049593620002269745, + 0.0684017539024353, + 0.060150932520627975, + 0.08451198041439056, + 0.012814555317163467, + 0.005495492834597826, + 0.006746881641447544, + 0.0035171587951481342, + 0.00557135371491313, + 0.005912405904382467, + 0.004114202689379454, + 0.007681094575673342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.03370063751935959 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048661816865205765 + ], + "opposite_logprob_tokens": [ + -3.048661708831787 + ], + "logprob_reward": -0.048661816865205765, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048661816865205765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20720906555652618, + 0.23651084303855896, + 0.1471327543258667, + 0.04295974597334862, + 0.04850069433450699, + 0.07903499901294708, + 0.030341995880007744, + 0.02525201253592968, + 0.11166249960660934, + 0.0713953971862793 + ], + "credit_map": [ + 0.06243898347020149, + 0.014399637468159199, + 0.014594761654734612, + 0.016700375825166702, + 0.010260793380439281, + 0.044622600078582764, + 0.02526608109474182, + 0.018925854936242104, + 0.014426988549530506, + 0.01622919738292694, + 0.015592656098306179, + 0.01810409501194954, + 0.018480351194739342, + 0.06861656904220581, + 0.03136628493666649, + 0.0536947175860405, + 0.03569263219833374, + 0.016896815970540047, + 0.014654784463346004, + 0.018199143931269646, + 0.013225789181888103, + 0.008849980309605598, + 0.01424973551183939, + 0.025363875553011894, + 0.004215971566736698, + 0.00478695472702384, + 0.005344218574464321, + 0.0046914550475776196, + 0.0059436531737446785, + 0.003781448118388653, + 0.006586759351193905, + 0.00760928587988019, + 0.008978940546512604, + 0.006981768179684877, + 0.005453493446111679, + 0.0060088192112743855, + 0.006167067680507898, + 0.004430355504155159, + 0.004485949873924255, + 0.005994305945932865, + 0.011419986374676228, + 0.008142168633639812, + 0.008495664224028587, + 0.012451313436031342, + 0.00910433940589428, + 0.006789733190089464, + 0.010422476567327976, + 0.012209322303533554, + 0.006973261013627052, + 0.0036156857386231422, + 0.0025244131684303284, + 0.0024625244550406933, + 0.002498553367331624, + 0.0025620278902351856, + 0.00440625986084342, + 0.005299271084368229, + 0.0037536504678428173, + 0.0020444057881832123, + 0.002783885458484292, + 0.002500450238585472, + 0.0033173684496432543, + 0.003939379937946796, + 0.0025864376220852137, + 0.00432643573731184, + 0.02031124383211136, + 0.009267839603126049, + 0.01060884352773428, + 0.008247754536569118, + 0.009034208953380585, + 0.019511517137289047, + 0.011139319278299809, + 0.023541778326034546, + 0.015419543720781803, + 0.008172545582056046, + 0.008252407424151897, + 0.006202310789376497, + 0.006580376531928778, + 0.009871544316411018, + 0.005305384751409292, + 0.011591292917728424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.048661816865205765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193302020430565 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193302020430565, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193302020430565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4523075222969055, + 0.035732582211494446, + 0.049819327890872955, + 0.0800345242023468, + 0.04541584476828575, + 0.044505827128887177, + 0.08118442445993423, + 0.034955866634845734, + 0.13239791989326477, + 0.04364613816142082 + ], + "credit_map": [ + 0.14791584014892578, + 0.023972462862730026, + 0.03366386890411377, + 0.02316424809396267, + 0.022003380581736565, + 0.12882009148597717, + 0.04745316132903099, + 0.025314485654234886, + 0.004408849403262138, + 0.003952076658606529, + 0.0026114792563021183, + 0.002493602456524968, + 0.0034392450470477343, + 0.0055145928636193275, + 0.005480076652020216, + 0.007832663133740425, + 0.00690218061208725, + 0.00419379398226738, + 0.005628546234220266, + 0.009944366291165352, + 0.00653597479686141, + 0.004556278232485056, + 0.004895389080047607, + 0.007162800990045071, + 0.008957252837717533, + 0.00764958281069994, + 0.010197482071816921, + 0.012802008539438248, + 0.008873715996742249, + 0.006434539798647165, + 0.011354170739650726, + 0.013765769079327583, + 0.005790672264993191, + 0.005968359299004078, + 0.005621835123747587, + 0.00762208690866828, + 0.005927767604589462, + 0.004667035304009914, + 0.004185908008366823, + 0.005632183980196714, + 0.006400216370820999, + 0.00378978019580245, + 0.005511065945029259, + 0.0076224361546337605, + 0.008012241683900356, + 0.0038906638510525227, + 0.004183679819107056, + 0.0050957477651536465, + 0.01396909449249506, + 0.007490043994039297, + 0.009013562463223934, + 0.007648938801139593, + 0.007357412949204445, + 0.008665000088512897, + 0.012492557987570763, + 0.014547819271683693, + 0.004333496559411287, + 0.002791119972243905, + 0.0035886478144675493, + 0.00427237106487155, + 0.0046174549497663975, + 0.006469195242971182, + 0.003763259155675769, + 0.005120320711284876, + 0.015993546694517136, + 0.009589043445885181, + 0.013157864101231098, + 0.01210160180926323, + 0.013922808691859245, + 0.02235126681625843, + 0.01821725256741047, + 0.027064546942710876, + 0.005626145284622908, + 0.0035529793240129948, + 0.004834027960896492, + 0.0051526413299143314, + 0.006268148776143789, + 0.0060606286861002445, + 0.005995154846459627, + 0.006156416144222021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.014193302020430565 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly rotating clockwise medium shot, filmed from a slightly upward perspective, captures a mythical phoenix. Each of its feathers appears to be forged from pure liquid gold, radiating a warm, dazzling glow. This light flows slowly through its wings like molten lava, while occasionally tiny golden sparks flake from the wings' edges, scattering like stardust against the deep, dark backdrop of cosmic nebulae. The phoenix is in the climax of a graceful movement, fully extending its massive, shimmering wings, nearly filling the entire frame. Its slender tail traces a brilliant arc of golden light behind it. Its body executes a balletic rotation in a steady and leisurely rhythm, its majestic, slender crown slightly raised, its blazing ruby eyes gazing into the distance. The entire scene is enveloped in an atmosphere of sacredness, tranquility, and awe. The dominant colors are brilliant gold and deep black, punctuated by faint, distant shimmers of purple and blue starlight in the background, evoking a mythical creation scene from the dawn of the universe.", + "seed": 8716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_2/index_6534_seed_8716_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03319674097001553, + "vlm_logprob": -0.03319674097001553, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.03319674097001553, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06088443472981453 + ], + "opposite_logprob_tokens": [ + -2.935884475708008 + ], + "logprob_reward": -0.06088443472981453, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06088443472981453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06952471286058426, + 0.09351984411478043, + 0.12605318427085876, + 0.0780475065112114, + 0.1302206665277481, + 0.10575515776872635, + 0.05194851756095886, + 0.11242011934518814, + 0.12388695776462555, + 0.10862329602241516 + ], + "credit_map": [ + 0.006955266930162907, + 0.006951718591153622, + 0.01191750168800354, + 0.005071769934147596, + 0.00478709489107132, + 0.014164933934807777, + 0.005570714827626944, + 0.014105713926255703, + 0.0087833721190691, + 0.009170780889689922, + 0.019662123173475266, + 0.01046889740973711, + 0.009313687682151794, + 0.013970918022096157, + 0.006186708342283964, + 0.015963353216648102, + 0.019328555092215538, + 0.009486839175224304, + 0.021489446982741356, + 0.010761551558971405, + 0.0042182025499641895, + 0.013558613136410713, + 0.009274766780436039, + 0.037935201078653336, + 0.006356756668537855, + 0.029211819171905518, + 0.014604013413190842, + 0.005655887071043253, + 0.00499061681330204, + 0.0070371320471167564, + 0.004637871868908405, + 0.005553417839109898, + 0.02892288565635681, + 0.015488263219594955, + 0.020854337140917778, + 0.007160597015172243, + 0.011384904384613037, + 0.01457926258444786, + 0.007067241705954075, + 0.024763191118836403, + 0.009736725129187107, + 0.009328925982117653, + 0.01395338587462902, + 0.01073444914072752, + 0.009132882580161095, + 0.016754083335399628, + 0.011726130731403828, + 0.024388570338487625, + 0.009016457013785839, + 0.006894180551171303, + 0.004228664096444845, + 0.004758886061608791, + 0.007279201410710812, + 0.0069593521766364574, + 0.005275252740830183, + 0.00753652211278677, + 0.01304015051573515, + 0.010531081818044186, + 0.013188440352678299, + 0.016964290291070938, + 0.01938864216208458, + 0.01444730069488287, + 0.013653759844601154, + 0.01120645273476839, + 0.017025236040353775, + 0.009042304009199142, + 0.010018355213105679, + 0.025415567681193352, + 0.013107451610267162, + 0.01646254025399685, + 0.019523240625858307, + 0.013292264193296432, + 0.014948302879929543, + 0.008173363283276558, + 0.018597790971398354, + 0.02418815717101097, + 0.009969661012291908, + 0.008343216963112354, + 0.014674649573862553, + 0.009728149510920048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.06088443472981453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012565843760967255 + ], + "opposite_logprob_tokens": [ + -4.387565612792969 + ], + "logprob_reward": -0.012565843760967255, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012565843760967255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049752119928598404, + 0.10126050561666489, + 0.10181253403425217, + 0.06522776186466217, + 0.3009100556373596, + 0.05379810929298401, + 0.0872245579957962, + 0.08584897220134735, + 0.09177426248788834, + 0.06239111348986626 + ], + "credit_map": [ + 0.004609114024788141, + 0.0037424848414957523, + 0.007771283388137817, + 0.0029864117968827486, + 0.004325837828218937, + 0.010912362486124039, + 0.0034801564179360867, + 0.01192447915673256, + 0.012736242264509201, + 0.011430779471993446, + 0.017406223341822624, + 0.008115398697555065, + 0.012279060669243336, + 0.015019901096820831, + 0.009393780492246151, + 0.01487912330776453, + 0.015605350024998188, + 0.00924587156623602, + 0.015556219033896923, + 0.012706668116152287, + 0.007396441884338856, + 0.010623934678733349, + 0.007742099463939667, + 0.022935954853892326, + 0.006918665021657944, + 0.023928163573145866, + 0.011469769291579723, + 0.005486910697072744, + 0.0035839660558849573, + 0.004977368284016848, + 0.003075245302170515, + 0.005787678062915802, + 0.07124770432710648, + 0.0379091314971447, + 0.04998817294836044, + 0.01804312691092491, + 0.02116047590970993, + 0.028498481959104538, + 0.015420259907841682, + 0.058642711490392685, + 0.004621242638677359, + 0.004955650307238102, + 0.006776421796530485, + 0.004919389262795448, + 0.0052905152551829815, + 0.007960866205394268, + 0.009266570210456848, + 0.010007459670305252, + 0.013436686247587204, + 0.007802235893905163, + 0.007385015022009611, + 0.009681891649961472, + 0.010519160889089108, + 0.013746040873229504, + 0.009154641069471836, + 0.015498886816203594, + 0.010351227596402168, + 0.006058810278773308, + 0.009004784747958183, + 0.014921397902071476, + 0.012606266885995865, + 0.011727357283234596, + 0.010465518571436405, + 0.01071360893547535, + 0.0162452831864357, + 0.00915549136698246, + 0.010328307747840881, + 0.011881547048687935, + 0.012421084567904472, + 0.0121085774153471, + 0.010468774475157261, + 0.00916520319879055, + 0.010929037816822529, + 0.005567781161516905, + 0.008287793956696987, + 0.012742775492370129, + 0.004948308225721121, + 0.006938385311514139, + 0.007380390074104071, + 0.005596645642071962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.012565843760967255 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329517900943756 + ], + "opposite_logprob_tokens": [ + -3.7732951641082764 + ], + "logprob_reward": -0.02329517900943756, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329517900943756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16787290573120117, + 0.07628405094146729, + 0.03985600173473358, + 0.09057172387838364, + 0.11036811023950577, + 0.08322731405496597, + 0.1473526507616043, + 0.10870900005102158, + 0.11821453273296356, + 0.05754369869828224 + ], + "credit_map": [ + 0.016927797347307205, + 0.022827869281172752, + 0.027029968798160553, + 0.014161254279315472, + 0.01345068495720625, + 0.027981797233223915, + 0.014841102063655853, + 0.030652422457933426, + 0.007901915349066257, + 0.00946386344730854, + 0.011386844329535961, + 0.009709913283586502, + 0.011393433436751366, + 0.008479833602905273, + 0.00739406980574131, + 0.010554177686572075, + 0.004865729715675116, + 0.00295620015822351, + 0.006360076367855072, + 0.004500397946685553, + 0.0038884577807039022, + 0.005210063885897398, + 0.002892474178224802, + 0.009182599373161793, + 0.00855189934372902, + 0.03140249475836754, + 0.014278390444815159, + 0.009918161667883396, + 0.005062254145741463, + 0.007808289490640163, + 0.006284067872911692, + 0.007266166619956493, + 0.007925056852400303, + 0.014471450820565224, + 0.016348090022802353, + 0.012110751122236252, + 0.0195928905159235, + 0.012315703555941582, + 0.014427928254008293, + 0.013176237232983112, + 0.007239730097353458, + 0.008751679211854935, + 0.012666272930800915, + 0.008031276986002922, + 0.010127449408173561, + 0.011750794015824795, + 0.0072909994050860405, + 0.017369117587804794, + 0.01657324656844139, + 0.012366535142064095, + 0.013104100711643696, + 0.015124785713851452, + 0.024047043174505234, + 0.021223057061433792, + 0.02197890728712082, + 0.02293497323989868, + 0.019802046939730644, + 0.008213112130761147, + 0.011947556398808956, + 0.013318813405930996, + 0.013266053050756454, + 0.01569536328315735, + 0.012561935000121593, + 0.013904121704399586, + 0.015224174596369267, + 0.007882189936935902, + 0.007967834360897541, + 0.03909362852573395, + 0.008649280294775963, + 0.009980933740735054, + 0.020589562132954597, + 0.008826926350593567, + 0.008169681765139103, + 0.005054504610598087, + 0.006529479753226042, + 0.007567461114376783, + 0.00963507778942585, + 0.006858809385448694, + 0.008043613284826279, + 0.005685068666934967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.02329517900943756 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865909367799759 + ], + "opposite_logprob_tokens": [ + -3.048659086227417 + ], + "logprob_reward": -0.04865909367799759, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865909367799759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39847448468208313, + 0.0504358634352684, + 0.05912650749087334, + 0.09806733578443527, + 0.01941394992172718, + 0.06297615170478821, + 0.09073838591575623, + 0.10475395619869232, + 0.05729358270764351, + 0.05871976539492607 + ], + "credit_map": [ + 0.02474163845181465, + 0.019399544224143028, + 0.08836478739976883, + 0.02093292959034443, + 0.02451656013727188, + 0.09637396782636642, + 0.017634475603699684, + 0.10651056468486786, + 0.007119228597730398, + 0.008594951592385769, + 0.007681631483137608, + 0.003962599206715822, + 0.0051553393714129925, + 0.006605051923543215, + 0.004080582410097122, + 0.0072364844381809235, + 0.011280657723546028, + 0.00466502970084548, + 0.009028393775224686, + 0.00532849645242095, + 0.0026365856174379587, + 0.006794799119234085, + 0.004537512548267841, + 0.014855027198791504, + 0.009282304905354977, + 0.040326330810785294, + 0.017853356897830963, + 0.005598117131739855, + 0.005548351909965277, + 0.007490435149520636, + 0.00484564621001482, + 0.0071228002198040485, + 0.0036227055825293064, + 0.0025398533325642347, + 0.0031018455047160387, + 0.0017650314839556813, + 0.0022546148393303156, + 0.001937433728016913, + 0.001440522144548595, + 0.0027519441209733486, + 0.005344562232494354, + 0.005090649705380201, + 0.008973880670964718, + 0.005670001730322838, + 0.007242436055094004, + 0.01202282402664423, + 0.00777796283364296, + 0.010853832587599754, + 0.015528419986367226, + 0.009481648914515972, + 0.005870717111974955, + 0.00869077630341053, + 0.010584264993667603, + 0.013128137215971947, + 0.011181146837770939, + 0.01627328246831894, + 0.01461314968764782, + 0.007540669292211533, + 0.012113582342863083, + 0.017638055607676506, + 0.013969868421554565, + 0.014623795635998249, + 0.011974317952990532, + 0.012280518189072609, + 0.009379290044307709, + 0.002523311646655202, + 0.007028176914900541, + 0.016524283215403557, + 0.0035006594844162464, + 0.006525907199829817, + 0.007657875772565603, + 0.004154081456363201, + 0.008415124379098415, + 0.0069128177128732204, + 0.008611483499407768, + 0.013931612484157085, + 0.005288419313728809, + 0.004934745375066996, + 0.006261104252189398, + 0.004364461172372103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.04865909367799759 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579153671860695 + ], + "opposite_logprob_tokens": [ + -3.8955790996551514 + ], + "logprob_reward": -0.020579153671860695, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579153671860695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19891522824764252, + 0.0903196632862091, + 0.09862757474184036, + 0.021941857412457466, + 0.056659404188394547, + 0.09179408848285675, + 0.11437089741230011, + 0.06399496644735336, + 0.07895420491695404, + 0.1844220906496048 + ], + "credit_map": [ + 0.013370887376368046, + 0.010975304059684277, + 0.03846590965986252, + 0.010494094341993332, + 0.015449811704456806, + 0.047439344227313995, + 0.013702345080673695, + 0.0490175262093544, + 0.01214767899364233, + 0.009705348871648312, + 0.015269795432686806, + 0.0076089720241725445, + 0.008942740969359875, + 0.013833235949277878, + 0.008035792969167233, + 0.014776092953979969, + 0.008131877519190311, + 0.00595133425667882, + 0.019849399104714394, + 0.01448585744947195, + 0.004745267331600189, + 0.008635836653411388, + 0.007445257622748613, + 0.029382750391960144, + 0.0027255816385149956, + 0.002731748390942812, + 0.0027318180073052645, + 0.0031610410660505295, + 0.0027904545422643423, + 0.0018570427782833576, + 0.0024237330071628094, + 0.0035204377491027117, + 0.009210070595145226, + 0.007017090450972319, + 0.008865712210536003, + 0.003957414533942938, + 0.005803954321891069, + 0.007284830790013075, + 0.004306796006858349, + 0.010213539004325867, + 0.008461380377411842, + 0.006591795943677425, + 0.014320623129606247, + 0.009225647896528244, + 0.009751925244927406, + 0.017082365229725838, + 0.009445429779589176, + 0.016914917156100273, + 0.016634400933980942, + 0.011803551577031612, + 0.00886526145040989, + 0.011165246367454529, + 0.011352092027664185, + 0.017960384488105774, + 0.015964049845933914, + 0.02062591351568699, + 0.008090286515653133, + 0.006895558442920446, + 0.006049216724932194, + 0.005703697446733713, + 0.013465437106788158, + 0.009067557752132416, + 0.00877827126532793, + 0.00594494491815567, + 0.011065108701586723, + 0.006319222040474415, + 0.00600010808557272, + 0.017951298505067825, + 0.007482384797185659, + 0.010441187769174576, + 0.012779048644006252, + 0.006915845908224583, + 0.029631925746798515, + 0.013015206903219223, + 0.025235556066036224, + 0.035653650760650635, + 0.013311534188687801, + 0.028394874185323715, + 0.02204233594238758, + 0.017137011513113976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.020579153671860695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly rotating clockwise medium shot, filmed from a slightly upward perspective, captures a mythical phoenix. Each of its feathers appears to be forged from pure liquid gold, radiating a warm, dazzling glow. This light flows slowly through its wings like molten lava, while occasionally tiny golden sparks flake from the wings' edges, scattering like stardust against the deep, dark backdrop of cosmic nebulae. The phoenix is in the climax of a graceful movement, fully extending its massive, shimmering wings, nearly filling the entire frame. Its slender tail traces a brilliant arc of golden light behind it. Its body executes a balletic rotation in a steady and leisurely rhythm, its majestic, slender crown slightly raised, its blazing ruby eyes gazing into the distance. The entire scene is enveloped in an atmosphere of sacredness, tranquility, and awe. The dominant colors are brilliant gold and deep black, punctuated by faint, distant shimmers of purple and blue starlight in the background, evoking a mythical creation scene from the dawn of the universe.", + "seed": 8717, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_2/index_6534_seed_8717_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02830226793885231, + "vlm_logprob": -0.02830226793885231, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.02830226793885231, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041327886283397675 + ], + "opposite_logprob_tokens": [ + -3.291327953338623 + ], + "logprob_reward": -0.041327886283397675, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.041327886283397675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06679671257734299, + 0.081849105656147, + 0.07297904789447784, + 0.10614824295043945, + 0.08572117239236832, + 0.07119932025671005, + 0.06220768019556999, + 0.20356228947639465, + 0.1971072554588318, + 0.05242921784520149 + ], + "credit_map": [ + 0.008409052155911922, + 0.009994559921324253, + 0.012564144097268581, + 0.010056539438664913, + 0.00541969807818532, + 0.0068473112769424915, + 0.007247328758239746, + 0.006258069537580013, + 0.009868117049336433, + 0.006807279773056507, + 0.011014916934072971, + 0.012120409868657589, + 0.0069090137258172035, + 0.009917602874338627, + 0.013914003036916256, + 0.011297752149403095, + 0.01033040788024664, + 0.00781921111047268, + 0.009943532757461071, + 0.005356156267225742, + 0.013755773194134235, + 0.00903767254203558, + 0.00845770351588726, + 0.008278589695692062, + 0.01941649615764618, + 0.009396038018167019, + 0.021023478358983994, + 0.011359000578522682, + 0.008975616656243801, + 0.011644146405160427, + 0.01230019610375166, + 0.01203326229006052, + 0.009366905316710472, + 0.010780776850879192, + 0.018184810876846313, + 0.022783750668168068, + 0.005805391818284988, + 0.00461011566221714, + 0.009262975305318832, + 0.004926443099975586, + 0.009737537242472172, + 0.00568490382283926, + 0.012883028946816921, + 0.015301770530641079, + 0.003841112833470106, + 0.005213720258325338, + 0.011800041422247887, + 0.006737199611961842, + 0.008736684918403625, + 0.0036218096502125263, + 0.008930916897952557, + 0.016886360943317413, + 0.003363063558936119, + 0.004761650692671537, + 0.00644779484719038, + 0.009459394961595535, + 0.03765860199928284, + 0.009579908102750778, + 0.02745719440281391, + 0.022906849160790443, + 0.02275238372385502, + 0.020774589851498604, + 0.038102149963378906, + 0.02433060109615326, + 0.03601503744721413, + 0.01907913014292717, + 0.02839963510632515, + 0.014704270288348198, + 0.013889982365071774, + 0.022674579173326492, + 0.03287845104932785, + 0.029466144740581512, + 0.006639058236032724, + 0.004141866695135832, + 0.007663961034268141, + 0.005694085732102394, + 0.0032211565412580967, + 0.009316796436905861, + 0.004764316603541374, + 0.010987969115376472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.041327886283397675 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812002018094063 + ], + "opposite_logprob_tokens": [ + -3.2881200313568115 + ], + "logprob_reward": -0.03812002018094063, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812002018094063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05744882673025131, + 0.07738008350133896, + 0.04057501256465912, + 0.04062323272228241, + 0.05941663682460785, + 0.07536305487155914, + 0.42869290709495544, + 0.09829369187355042, + 0.07014387845993042, + 0.052062708884477615 + ], + "credit_map": [ + 0.00606395723298192, + 0.006206012796610594, + 0.016755666583776474, + 0.009430230595171452, + 0.0036928714253008366, + 0.004136298783123493, + 0.006354554556310177, + 0.004809232894331217, + 0.00954427570104599, + 0.007488153409212828, + 0.007761151995509863, + 0.0162886381149292, + 0.005800684913992882, + 0.010125610046088696, + 0.011449352838099003, + 0.008922207169234753, + 0.013350250199437141, + 0.0021868888288736343, + 0.012751886621117592, + 0.002636171178892255, + 0.0016136111225932837, + 0.002078706631436944, + 0.0031682418193668127, + 0.0027892510406672955, + 0.009380394592881203, + 0.002661449834704399, + 0.009533709846436977, + 0.0044057853519916534, + 0.0033961173612624407, + 0.0032823735382407904, + 0.004510981496423483, + 0.0034524211660027504, + 0.006715207360684872, + 0.010443514212965965, + 0.009065977297723293, + 0.009482825174927711, + 0.00519423745572567, + 0.004553871229290962, + 0.0090031111612916, + 0.004957889206707478, + 0.0063851140439510345, + 0.0043056318536400795, + 0.01429462619125843, + 0.01963069848716259, + 0.004412458743900061, + 0.004909881390631199, + 0.011809280142188072, + 0.009615357033908367, + 0.10002045333385468, + 0.020202798768877983, + 0.0455474779009819, + 0.16759313642978668, + 0.011938463896512985, + 0.015633689239621162, + 0.03810281306505203, + 0.029654068872332573, + 0.01982247643172741, + 0.007087944075465202, + 0.01871477998793125, + 0.010710746049880981, + 0.00862993486225605, + 0.011187326163053513, + 0.011249887757003307, + 0.010890593752264977, + 0.012558357790112495, + 0.007460727822035551, + 0.009291165508329868, + 0.006446965038776398, + 0.0070342510007321835, + 0.01039476040750742, + 0.00975518487393856, + 0.00720245623961091, + 0.0071376170963048935, + 0.004279084037989378, + 0.008183851838111877, + 0.004148124251514673, + 0.003958288114517927, + 0.009213482961058617, + 0.002699321834370494, + 0.012442938052117825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.03812002018094063 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197767712175846 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197767712175846, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197767712175846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047049518674612045, + 0.03396158292889595, + 0.15605786442756653, + 0.06843192130327225, + 0.08094944059848785, + 0.12208051979541779, + 0.27875223755836487, + 0.09402952343225479, + 0.0758739486336708, + 0.042813464999198914 + ], + "credit_map": [ + 0.007939597591757774, + 0.005354092922061682, + 0.008970800787210464, + 0.006649330724030733, + 0.003433611011132598, + 0.004198161419481039, + 0.006353630684316158, + 0.00415028864517808, + 0.005222994834184647, + 0.003172541968524456, + 0.008077265694737434, + 0.004538089968264103, + 0.0023728024680167437, + 0.0028807406779378653, + 0.00425025774165988, + 0.00344688817858696, + 0.04360928386449814, + 0.011457913555204868, + 0.0425456166267395, + 0.009166394360363483, + 0.020547784864902496, + 0.011393359862267971, + 0.010333743877708912, + 0.007003764156252146, + 0.012350128963589668, + 0.006354583892971277, + 0.012184754945337772, + 0.00902977492660284, + 0.005528070032596588, + 0.00645214831456542, + 0.009499611333012581, + 0.0070328484289348125, + 0.007469370029866695, + 0.013900079764425755, + 0.014132148586213589, + 0.016198350116610527, + 0.0050926306284964085, + 0.007085662335157394, + 0.009541859850287437, + 0.007529325783252716, + 0.014464366249740124, + 0.009637451730668545, + 0.018628116697072983, + 0.024793613702058792, + 0.009802920743823051, + 0.011485124938189983, + 0.017720062285661697, + 0.01554886344820261, + 0.06376305967569351, + 0.01408469770103693, + 0.03325207158923149, + 0.11074063181877136, + 0.00799605157226324, + 0.0082679009065032, + 0.022015245631337166, + 0.018632570281624794, + 0.01745341345667839, + 0.00861282367259264, + 0.014864826574921608, + 0.01012575626373291, + 0.009622598066926003, + 0.01178685761988163, + 0.008229561150074005, + 0.013333681039512157, + 0.015104095451533794, + 0.006440815050154924, + 0.012596622109413147, + 0.00745511194691062, + 0.004874549340456724, + 0.01388291735202074, + 0.007526344619691372, + 0.007993492297828197, + 0.005858249496668577, + 0.0038148085586726665, + 0.005870818626135588, + 0.0042768740095198154, + 0.003736421000212431, + 0.007276394870132208, + 0.0036576311103999615, + 0.008322269655764103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.014197767712175846 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977869238704443 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.00977869238704443, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977869238704443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061321455985307693, + 0.030110593885183334, + 0.06006588786840439, + 0.05542512610554695, + 0.3015500009059906, + 0.03569873794913292, + 0.31611305475234985, + 0.03703318163752556, + 0.04463152214884758, + 0.05805051326751709 + ], + "credit_map": [ + 0.008236969821155071, + 0.008727253414690495, + 0.011368084698915482, + 0.008230336010456085, + 0.0038895998150110245, + 0.007568397559225559, + 0.007386299781501293, + 0.005914510227739811, + 0.0037852362729609013, + 0.0027623865753412247, + 0.003812777576968074, + 0.006006340496242046, + 0.002386294538155198, + 0.002787133911624551, + 0.005289048887789249, + 0.003281370038166642, + 0.010418371297419071, + 0.006321306806057692, + 0.009091902524232864, + 0.004123535472899675, + 0.011789989657700062, + 0.006779706571251154, + 0.0069497125223279, + 0.004591362550854683, + 0.017084702849388123, + 0.0032394330482929945, + 0.017060235142707825, + 0.003903021803125739, + 0.003619737457484007, + 0.0036721276119351387, + 0.0036528403870761395, + 0.0031930184923112392, + 0.030698837712407112, + 0.02103463187813759, + 0.07539539039134979, + 0.10466644912958145, + 0.013229342177510262, + 0.01067716721445322, + 0.030633769929409027, + 0.015214379876852036, + 0.004331192467361689, + 0.003934753593057394, + 0.006737785879522562, + 0.006721829995512962, + 0.0025951473508030176, + 0.0027758777141571045, + 0.004712950903922319, + 0.003889196552336216, + 0.0688001960515976, + 0.013311509974300861, + 0.031000036746263504, + 0.1096121147274971, + 0.009648876264691353, + 0.011699113994836807, + 0.030199425294995308, + 0.04184173792600632, + 0.006943440530449152, + 0.002098464872688055, + 0.005141264293342829, + 0.004130603279918432, + 0.0037062205374240875, + 0.003712671110406518, + 0.006524378899484873, + 0.004776135087013245, + 0.0070804753340780735, + 0.003729458898305893, + 0.007278531789779663, + 0.004379723686724901, + 0.0037699230015277863, + 0.006968400441110134, + 0.00546817434951663, + 0.005956834647804499, + 0.006504890974611044, + 0.004240889567881823, + 0.00852450542151928, + 0.00622357614338398, + 0.004335446283221245, + 0.010696766898036003, + 0.004638617392629385, + 0.012885818257927895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.00977869238704443 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808697313070297 + ], + "opposite_logprob_tokens": [ + -3.2880868911743164 + ], + "logprob_reward": -0.03808697313070297, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808697313070297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12511707842350006, + 0.02879089117050171, + 0.36521655321121216, + 0.10416048020124435, + 0.1269129514694214, + 0.06960076838731766, + 0.05984224006533623, + 0.045467063784599304, + 0.028875697404146194, + 0.04601623862981796 + ], + "credit_map": [ + 0.0086995093151927, + 0.01472532469779253, + 0.03439478576183319, + 0.020774805918335915, + 0.008205576799809933, + 0.012920527718961239, + 0.0143323615193367, + 0.011064194142818451, + 0.003094273852184415, + 0.0027690136339515448, + 0.003149911528453231, + 0.004276043735444546, + 0.0030250202398747206, + 0.0030310091096907854, + 0.004653159528970718, + 0.0047924574464559555, + 0.14249394834041595, + 0.016716867685317993, + 0.13871684670448303, + 0.009601340629160404, + 0.01506865955889225, + 0.015835164114832878, + 0.01363828219473362, + 0.013145451433956623, + 0.017497096210718155, + 0.00884170550853014, + 0.02350359596312046, + 0.01056410651654005, + 0.010000879876315594, + 0.010982861742377281, + 0.013493574224412441, + 0.009276661090552807, + 0.012863798066973686, + 0.013034134171903133, + 0.027098024263978004, + 0.03609885647892952, + 0.007644194643944502, + 0.007695287000387907, + 0.013746818527579308, + 0.008731838315725327, + 0.009155812673270702, + 0.006243772804737091, + 0.01045637670904398, + 0.015956824645400047, + 0.004899242892861366, + 0.006458746735006571, + 0.008308742195367813, + 0.0081212492659688, + 0.006857661530375481, + 0.003978326451033354, + 0.012599579058587551, + 0.017271244898438454, + 0.002666222397238016, + 0.004458644427359104, + 0.005290347617119551, + 0.006720214616507292, + 0.009446321055293083, + 0.0030886742752045393, + 0.00837477296590805, + 0.005186615511775017, + 0.004962595645338297, + 0.005571136716753244, + 0.00415705656632781, + 0.004679891746491194, + 0.006392435636371374, + 0.0030277231708168983, + 0.0036188974045217037, + 0.003653709776699543, + 0.001995408209040761, + 0.0041665672324597836, + 0.0029549177270382643, + 0.0030660394113510847, + 0.004801868461072445, + 0.003296379465609789, + 0.007951557636260986, + 0.004342276602983475, + 0.005049717612564564, + 0.008463681675493717, + 0.002631786046549678, + 0.009478970430791378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.03808697313070297 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly rotating clockwise medium shot, filmed from a slightly upward perspective, captures a mythical phoenix. Each of its feathers appears to be forged from pure liquid gold, radiating a warm, dazzling glow. This light flows slowly through its wings like molten lava, while occasionally tiny golden sparks flake from the wings' edges, scattering like stardust against the deep, dark backdrop of cosmic nebulae. The phoenix is in the climax of a graceful movement, fully extending its massive, shimmering wings, nearly filling the entire frame. Its slender tail traces a brilliant arc of golden light behind it. Its body executes a balletic rotation in a steady and leisurely rhythm, its majestic, slender crown slightly raised, its blazing ruby eyes gazing into the distance. The entire scene is enveloped in an atmosphere of sacredness, tranquility, and awe. The dominant colors are brilliant gold and deep black, punctuated by faint, distant shimmers of purple and blue starlight in the background, evoking a mythical creation scene from the dawn of the universe.", + "seed": 8718, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_2/index_6534_seed_8718_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02899923175573349, + "vlm_logprob": -0.02899923175573349, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.02899923175573349, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048285678029060364 + ], + "opposite_logprob_tokens": [ + -3.173285722732544 + ], + "logprob_reward": -0.048285678029060364, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.048285678029060364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23540790379047394, + 0.1394476592540741, + 0.053981974720954895, + 0.06435756385326385, + 0.14059385657310486, + 0.07290538400411606, + 0.08485645800828934, + 0.06561410427093506, + 0.0829799547791481, + 0.059855200350284576 + ], + "credit_map": [ + 0.04852356016635895, + 0.03203055262565613, + 0.037862714380025864, + 0.02237219363451004, + 0.017216410487890244, + 0.024902548640966415, + 0.019252575933933258, + 0.03324730321764946, + 0.01614045910537243, + 0.012043769471347332, + 0.024980248883366585, + 0.01726333238184452, + 0.014852759428322315, + 0.013119485229253769, + 0.018651708960533142, + 0.022395869717001915, + 0.004457477014511824, + 0.0054520112462341785, + 0.01132150087505579, + 0.0052263168618083, + 0.0060311369597911835, + 0.005093165673315525, + 0.006635860074311495, + 0.009764501824975014, + 0.00548505736514926, + 0.005705793853849173, + 0.010979502461850643, + 0.007017186377197504, + 0.006032878067344427, + 0.00789220817387104, + 0.010181971825659275, + 0.01106295920908451, + 0.022796446457505226, + 0.014224573969841003, + 0.020539646968245506, + 0.015104887075722218, + 0.012682232074439526, + 0.018118444830179214, + 0.009847036562860012, + 0.027280572801828384, + 0.01465565711259842, + 0.007764431647956371, + 0.01506793312728405, + 0.006055947858840227, + 0.007071642205119133, + 0.008681122213602066, + 0.00494739506393671, + 0.008661244064569473, + 0.00940687209367752, + 0.007325668353587389, + 0.012117777019739151, + 0.007184693124145269, + 0.007288127671927214, + 0.018092766404151917, + 0.006802245508879423, + 0.016638295724987984, + 0.010503928177058697, + 0.005974888801574707, + 0.011250444687902927, + 0.004929929506033659, + 0.0048163472674787045, + 0.012780849821865559, + 0.005992485675960779, + 0.009365217760205269, + 0.008372333832085133, + 0.010081471875309944, + 0.011744542047381401, + 0.009188462048768997, + 0.006145081482827663, + 0.011014241725206375, + 0.008737567812204361, + 0.017696242779493332, + 0.00882701575756073, + 0.006862207781523466, + 0.007604957092553377, + 0.007961998693645, + 0.0073005398735404015, + 0.0074075148440897465, + 0.007198626175522804, + 0.006692331749945879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.048285678029060364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160874892026186 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.0160874892026186, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160874892026186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16108009219169617, + 0.12891823053359985, + 0.08867009729146957, + 0.08286825567483902, + 0.05040513351559639, + 0.04750584065914154, + 0.06804172694683075, + 0.07866664230823517, + 0.24778318405151367, + 0.046060848981142044 + ], + "credit_map": [ + 0.01586136966943741, + 0.02046745829284191, + 0.01805863343179226, + 0.0142858000472188, + 0.013615686446428299, + 0.021084101870656013, + 0.01981109008193016, + 0.03789594769477844, + 0.011793616227805614, + 0.013583226129412651, + 0.018311135470867157, + 0.01750539056956768, + 0.014165686443448067, + 0.011523661203682423, + 0.0207296684384346, + 0.021305840462446213, + 0.007669388782233, + 0.00792237464338541, + 0.018499132245779037, + 0.006624853704124689, + 0.010438908822834492, + 0.013409429229795933, + 0.009853425435721874, + 0.014252577908337116, + 0.008412952534854412, + 0.00837929267436266, + 0.016542913392186165, + 0.007309797685593367, + 0.00974776316434145, + 0.009709243662655354, + 0.008406580425798893, + 0.014359711669385433, + 0.0065286606550216675, + 0.003845551051199436, + 0.006760071497410536, + 0.004935864824801683, + 0.006835196167230606, + 0.007702950853854418, + 0.005808355286717415, + 0.007988483645021915, + 0.005122671835124493, + 0.0037857317365705967, + 0.008640630170702934, + 0.0050763594917953014, + 0.006458767224103212, + 0.006093264557421207, + 0.004412513226270676, + 0.007915904745459557, + 0.0061674658209085464, + 0.006452330853790045, + 0.009199999272823334, + 0.006624074187129736, + 0.011824076063930988, + 0.011897836811840534, + 0.006199592258781195, + 0.009676351211965084, + 0.008144557476043701, + 0.0049531166441738605, + 0.014213123358786106, + 0.006791343912482262, + 0.005855001974850893, + 0.008013611659407616, + 0.011063411831855774, + 0.019632471725344658, + 0.03516189008951187, + 0.02489287778735161, + 0.032656002789735794, + 0.03795061260461807, + 0.018060695379972458, + 0.034861598163843155, + 0.022780992090702057, + 0.04141851142048836, + 0.007549919653683901, + 0.005885590333491564, + 0.004688940476626158, + 0.004666716326028109, + 0.007936492562294006, + 0.005821099504828453, + 0.004761677235364914, + 0.004750409163534641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0160874892026186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635025978088379 + ], + "opposite_logprob_tokens": [ + -3.651350259780884 + ], + "logprob_reward": -0.02635025978088379, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635025978088379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2296903431415558, + 0.048523105680942535, + 0.047428954392671585, + 0.08359747380018234, + 0.03987564891576767, + 0.050138842314481735, + 0.18859457969665527, + 0.07527130097150803, + 0.19150198996067047, + 0.045377761125564575 + ], + "credit_map": [ + 0.042892955243587494, + 0.022741178050637245, + 0.03164268285036087, + 0.02210584655404091, + 0.01914406381547451, + 0.023035772144794464, + 0.021573131904006004, + 0.04655471071600914, + 0.004349717870354652, + 0.004443317651748657, + 0.007818217389285564, + 0.005613618064671755, + 0.0035203055012971163, + 0.00343515002168715, + 0.00683213584125042, + 0.012510642409324646, + 0.003544120816513896, + 0.0033233643043786287, + 0.009514521807432175, + 0.0039902059361338615, + 0.0075865620747208595, + 0.004620544146746397, + 0.004883502144366503, + 0.00996613223105669, + 0.007539020385593176, + 0.007335671689361334, + 0.01797017827630043, + 0.007175942417234182, + 0.005745834670960903, + 0.011287178844213486, + 0.009604962542653084, + 0.016938690096139908, + 0.004508429206907749, + 0.003447331255301833, + 0.005373014137148857, + 0.004246656317263842, + 0.004705183207988739, + 0.005611649248749018, + 0.00466727651655674, + 0.007316108327358961, + 0.0054136780090630054, + 0.004472250118851662, + 0.00779713923111558, + 0.00538245914503932, + 0.007202060427516699, + 0.008341842330992222, + 0.004102086182683706, + 0.007427326403558254, + 0.01925894059240818, + 0.017541727051138878, + 0.021865371614694595, + 0.012819487601518631, + 0.018084367737174034, + 0.04399055242538452, + 0.017685385420918465, + 0.03734876215457916, + 0.007390304934233427, + 0.006693263538181782, + 0.013798810541629791, + 0.007334973197430372, + 0.006536389235407114, + 0.008866638876497746, + 0.008905054070055485, + 0.015745867043733597, + 0.027699651196599007, + 0.02039874903857708, + 0.020648805424571037, + 0.03429773449897766, + 0.017910268157720566, + 0.027625951915979385, + 0.017425747588276863, + 0.02549508586525917, + 0.007675169967114925, + 0.005276903975754976, + 0.005131365731358528, + 0.005751048214733601, + 0.004465390462428331, + 0.006031875032931566, + 0.005394686944782734, + 0.0056513226591050625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.02635025978088379 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033695563673973083 + ], + "opposite_logprob_tokens": [ + -3.4086954593658447 + ], + "logprob_reward": -0.033695563673973083, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033695563673973083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2751629948616028, + 0.5868984460830688, + 0.011810041964054108, + 0.011299473233520985, + 0.018924584612250328, + 0.012247991748154163, + 0.010371951386332512, + 0.008100156672298908, + 0.042370837181806564, + 0.022813621908426285 + ], + "credit_map": [ + 0.03457910567522049, + 0.03880590200424194, + 0.03415021672844887, + 0.029761606827378273, + 0.019810182973742485, + 0.029469206929206848, + 0.044905636459589005, + 0.04368114471435547, + 0.04582791030406952, + 0.04122009873390198, + 0.102873794734478, + 0.065149687230587, + 0.06969171017408371, + 0.04762590676546097, + 0.07494445145130157, + 0.13956481218338013, + 0.0010733386734500527, + 0.001254489878192544, + 0.0021464836318045855, + 0.0010539263021200895, + 0.001492533483542502, + 0.0021457106340676546, + 0.001022315351292491, + 0.0016212441259995103, + 0.001557462615892291, + 0.000893550633918494, + 0.0015799106331542134, + 0.0007795862038619816, + 0.0016697229584679008, + 0.001720881788060069, + 0.0013234768994152546, + 0.0017748806858435273, + 0.0035017170011997223, + 0.0019523488590493798, + 0.0026545522268861532, + 0.002010483294725418, + 0.0020117571111768484, + 0.0023633947130292654, + 0.0013575477059930563, + 0.0030727824196219444, + 0.0014776232419535518, + 0.0010580612579360604, + 0.0020995596423745155, + 0.001277987495996058, + 0.0015093275578692555, + 0.0020995079539716244, + 0.0012041220907121897, + 0.0015218015760183334, + 0.0011637230636551976, + 0.0009480503504164517, + 0.0012670583091676235, + 0.0010904044611379504, + 0.0014494841452687979, + 0.001639731926843524, + 0.0008879938977770507, + 0.001925505232065916, + 0.001153715536929667, + 0.0007218719110824168, + 0.0011159736895933747, + 0.0005758641054853797, + 0.0007326157647185028, + 0.0015931760426610708, + 0.0006933899130672216, + 0.0015135500580072403, + 0.004062959924340248, + 0.004821807146072388, + 0.006160310935229063, + 0.00449750991538167, + 0.003006843151524663, + 0.005839538294821978, + 0.004475034773349762, + 0.009506834670901299, + 0.0033603045158088207, + 0.0025371031370013952, + 0.0032861477229744196, + 0.0024060290306806564, + 0.0028883316554129124, + 0.002856971463188529, + 0.002265987219288945, + 0.00321274739690125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.033695563673973083 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with feathers made of pure liquid gold?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the phoenix shown in the climax of a graceful movement with its massive wings fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the phoenix's body execute a slow, steady, clockwise rotation while its tail traces a brilliant arc of golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep, dark cosmic nebula with faint distant shimmers of purple and blue starlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a medium shot filmed from a slightly upward perspective where the phoenix nearly fills the entire frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577168092131615 + ], + "opposite_logprob_tokens": [ + -3.8955771923065186 + ], + "logprob_reward": -0.020577168092131615, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577168092131615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15623202919960022, + 0.2850136458873749, + 0.06283789128065109, + 0.08823244273662567, + 0.05926452949643135, + 0.10096298903226852, + 0.07705768942832947, + 0.06248191371560097, + 0.07457348704338074, + 0.03334341198205948 + ], + "credit_map": [ + 0.02481289952993393, + 0.01678224466741085, + 0.020005958154797554, + 0.01629205048084259, + 0.013994008302688599, + 0.01842031255364418, + 0.01597079634666443, + 0.02995375543832779, + 0.025467921048402786, + 0.019287176430225372, + 0.04669048637151718, + 0.027309395372867584, + 0.02886601723730564, + 0.0197432991117239, + 0.0335162952542305, + 0.08413303643465042, + 0.0042925490997731686, + 0.004998208489269018, + 0.01411344949156046, + 0.005707560107111931, + 0.008798694238066673, + 0.006234756205230951, + 0.005711590405553579, + 0.012981085106730461, + 0.0072085377760231495, + 0.008808942511677742, + 0.016608379781246185, + 0.009062484838068485, + 0.008927545510232449, + 0.011771798133850098, + 0.010412401519715786, + 0.015432353131473064, + 0.006656009703874588, + 0.0035203457809984684, + 0.009305254556238651, + 0.006271779537200928, + 0.005921201780438423, + 0.008993399329483509, + 0.006807051599025726, + 0.011789488606154919, + 0.013949014246463776, + 0.010047530755400658, + 0.019297994673252106, + 0.007232652045786381, + 0.014720981009304523, + 0.014390389434993267, + 0.007906784303486347, + 0.013417646288871765, + 0.006251749116927385, + 0.006238627713173628, + 0.009416112676262856, + 0.0068758041597902775, + 0.005839616991579533, + 0.01821465604007244, + 0.006585380062460899, + 0.01763574220240116, + 0.007911299355328083, + 0.005646722391247749, + 0.00963603239506483, + 0.005698861554265022, + 0.0067437319085001945, + 0.008728652261197567, + 0.007413512095808983, + 0.010703103616833687, + 0.006049364805221558, + 0.006786237936466932, + 0.008804718032479286, + 0.011583365499973297, + 0.013014591298997402, + 0.009321410208940506, + 0.009101198054850101, + 0.009912604466080666, + 0.006166406907141209, + 0.005144525784999132, + 0.004967082291841507, + 0.0032965412829071283, + 0.002376904245465994, + 0.004884242545813322, + 0.0034045048523694277, + 0.0031032038386911154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.020577168092131615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_3/index_tensor([2740, 2740])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_3/index_tensor([2740, 2740])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ea3a3e2d0a3542beb445cc6d5ada70c68e0e7a6e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_3/index_tensor([2740, 2740])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, long shot, rendered with the exquisite detail and majestic grandeur of classical Chinese art, captures a pivotal moment of conflict within a vast, dimly lit hall. Massive, dark-lacquered wooden pillars tower over the invisible, high ceiling, while golden light pours through ornate, high windows, permeating the thick atmosphere. At the center of this vast space, a large, heavy Chinese chess board crafted from polished, dark rosewood rests on a low table. The camera zooms in on an ebony chariot, its surface a deep, luminous black, as it violently crashes into an ornately carved ivory commander in breathtaking slow motion. The commander, crafted from aged ivory with a faint yellow patina and adorned with miniature dragon motifs around its base, is instantly launched upward and backward from the board. It tumbles chaotically through a bright beam of light, the crimson ink of the central figure momentarily flashing towards the camera. At the point of impact, a tiny cloud of ivory dust and microscopic debris explodes outward, catching the light like a fleeting, shimmering cloud, while the heavy slabs themselves visibly shimmer, emphasizing the silent, resonant power of the conflict and creating a dramatic spectacle of strategic endgame and controlled violence.", + "seed": 7262, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_3/index_2740_seed_7262_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1777670193463564, + "vlm_logprob": -0.1777670193463564, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.1777670193463564, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010492635890841484 + ], + "opposite_logprob_tokens": [ + -5.26049280166626 + ], + "logprob_reward": -0.010492635890841484, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.010492635890841484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.280758261680603, + 0.18193623423576355, + 0.07829516381025314, + 0.05195363610982895, + 0.03558526188135147, + 0.04564743861556053, + 0.11261576414108276, + 0.04825359955430031, + 0.038698628544807434, + 0.12625594437122345 + ], + "credit_map": [ + 0.05037879943847656, + 0.028678767383098602, + 0.04336137697100639, + 0.033973369747400284, + 0.028431223705410957, + 0.028708534315228462, + 0.03186916187405586, + 0.0353570431470871, + 0.026728399097919464, + 0.024644680321216583, + 0.024501383304595947, + 0.02527589350938797, + 0.02223045378923416, + 0.02236209250986576, + 0.018419766798615456, + 0.0177735798060894, + 0.011963558383286, + 0.011876695789396763, + 0.0101419473066926, + 0.00689685670658946, + 0.010171666741371155, + 0.010491007007658482, + 0.0072809974662959576, + 0.00947243720293045, + 0.006327319424599409, + 0.0054484773427248, + 0.008304610848426819, + 0.00768195278942585, + 0.006491250358521938, + 0.004917527083307505, + 0.008417530916631222, + 0.0043649692088365555, + 0.00573380384594202, + 0.0044143954291939735, + 0.005641964729875326, + 0.004611608572304249, + 0.0033781288657337427, + 0.003476747777312994, + 0.00485537713393569, + 0.00347323645837605, + 0.008208968676626682, + 0.004756394308060408, + 0.004593047313392162, + 0.006117764860391617, + 0.005051882937550545, + 0.006353866774588823, + 0.005752679891884327, + 0.004812838975340128, + 0.018262086436152458, + 0.021445276215672493, + 0.009675086475908756, + 0.01521680410951376, + 0.011187683790922165, + 0.008900897577404976, + 0.015297246165573597, + 0.012630697339773178, + 0.007881366647779942, + 0.0056053344160318375, + 0.005886588711291552, + 0.01010947860777378, + 0.0035305111669003963, + 0.002656801138073206, + 0.006122006103396416, + 0.006461513694375753, + 0.006139538250863552, + 0.004865953698754311, + 0.007007936481386423, + 0.004721609875559807, + 0.002862262539565563, + 0.0018724714173004031, + 0.006095454562455416, + 0.005133406724780798, + 0.044241055846214294, + 0.035181768238544464, + 0.010255699045956135, + 0.008476952090859413, + 0.005123927723616362, + 0.0026443100068718195, + 0.014697563834488392, + 0.005634688772261143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.010492635890841484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160874892026186 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.0160874892026186, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160874892026186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10885942727327347, + 0.0775497630238533, + 0.08467467129230499, + 0.03302556648850441, + 0.04361555352807045, + 0.07882663607597351, + 0.0851457342505455, + 0.10093996673822403, + 0.1179857850074768, + 0.26937687397003174 + ], + "credit_map": [ + 0.015437365509569645, + 0.011840919964015484, + 0.01791449449956417, + 0.01332471240311861, + 0.014057902619242668, + 0.007121190894395113, + 0.015884872525930405, + 0.013277966529130936, + 0.012094088830053806, + 0.01080482266843319, + 0.010996046476066113, + 0.013221499510109425, + 0.006538665387779474, + 0.006198231130838394, + 0.007568902801722288, + 0.010127505287528038, + 0.012017169035971165, + 0.01263060886412859, + 0.00896188523620367, + 0.012736334465444088, + 0.01112634688615799, + 0.008537368848919868, + 0.00992274098098278, + 0.008742215111851692, + 0.004900976549834013, + 0.004076089709997177, + 0.005291847512125969, + 0.004185624420642853, + 0.0034951104316860437, + 0.0033121530432254076, + 0.004456955473870039, + 0.0033068086486309767, + 0.006622618529945612, + 0.00593823567032814, + 0.009069662541151047, + 0.006075981073081493, + 0.0032733865082263947, + 0.003421785542741418, + 0.004893881268799305, + 0.00431999983265996, + 0.013531554490327835, + 0.0053787631914019585, + 0.008327960036695004, + 0.017603835090994835, + 0.008248185738921165, + 0.009289230220019817, + 0.009662140160799026, + 0.006784970872104168, + 0.016614601016044617, + 0.00994073785841465, + 0.007641374133527279, + 0.018767330795526505, + 0.005339490715414286, + 0.004001529887318611, + 0.015548217110335827, + 0.007292451802641153, + 0.015516236424446106, + 0.010336239822208881, + 0.011351235210895538, + 0.020876241847872734, + 0.005796751938760281, + 0.00497486162930727, + 0.021016744896769524, + 0.01107165589928627, + 0.03214108943939209, + 0.01566135324537754, + 0.014969171956181526, + 0.010280782356858253, + 0.006125914864242077, + 0.004179899115115404, + 0.025120152160525322, + 0.009507425129413605, + 0.08202730119228363, + 0.09962298721075058, + 0.019960811361670494, + 0.015263917855918407, + 0.010067187249660492, + 0.0057107736356556416, + 0.0264422744512558, + 0.010281645692884922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0160874892026186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263509564101696 + ], + "opposite_logprob_tokens": [ + -3.651350975036621 + ], + "logprob_reward": -0.0263509564101696, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263509564101696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35419341921806335, + 0.09766296297311783, + 0.2491072118282318, + 0.04050856828689575, + 0.06212735176086426, + 0.019729597494006157, + 0.04989601671695709, + 0.04274198040366173, + 0.02569103054702282, + 0.05834184214472771 + ], + "credit_map": [ + 0.055356621742248535, + 0.04113040491938591, + 0.0493328720331192, + 0.0384846031665802, + 0.05185510218143463, + 0.04338134825229645, + 0.04201683774590492, + 0.03263562172651291, + 0.018535366281867027, + 0.015798965469002724, + 0.012453858740627766, + 0.017461346462368965, + 0.00665700389072299, + 0.008225983008742332, + 0.008965650573372841, + 0.009564791806042194, + 0.02192750573158264, + 0.06367644667625427, + 0.03149513900279999, + 0.03813571855425835, + 0.028699245303869247, + 0.015502613969147205, + 0.031214352697134018, + 0.01845618709921837, + 0.0050836605951189995, + 0.005121997091919184, + 0.004540873225778341, + 0.004979691002517939, + 0.004105652216821909, + 0.0029449581634253263, + 0.008778475224971771, + 0.004953262396156788, + 0.009718059562146664, + 0.007928239181637764, + 0.015801183879375458, + 0.007962968200445175, + 0.004418381489813328, + 0.004290884826332331, + 0.006390690803527832, + 0.005616942886263132, + 0.002961850492283702, + 0.0020234989933669567, + 0.0022152455057948828, + 0.0035737515427172184, + 0.0014121143613010645, + 0.0013362244935706258, + 0.00386230181902647, + 0.002344610868021846, + 0.008852503262460232, + 0.0069269235245883465, + 0.005002585239708424, + 0.009755494073033333, + 0.003961884416639805, + 0.0037978014443069696, + 0.007336051668971777, + 0.004262774251401424, + 0.007037511095404625, + 0.00612490763887763, + 0.004332251381129026, + 0.008234995417296886, + 0.0032100295647978783, + 0.0023929327726364136, + 0.007461973000317812, + 0.003947383724153042, + 0.004334698431193829, + 0.0035201923456043005, + 0.0036880592815577984, + 0.003908834420144558, + 0.0023497857619076967, + 0.0013838630402460694, + 0.003958745393902063, + 0.002546851523220539, + 0.013440804556012154, + 0.019172141328454018, + 0.005209747236222029, + 0.005011440254747868, + 0.00270354887470603, + 0.001558826887048781, + 0.007565807085484266, + 0.0036795230116695166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0263509564101696 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778810665011406 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778810665011406, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778810665011406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16313719749450684, + 0.16992610692977905, + 0.14243964850902557, + 0.02705894038081169, + 0.09690911322832108, + 0.09797587245702744, + 0.054365336894989014, + 0.05132941156625748, + 0.0329042449593544, + 0.16395416855812073 + ], + "credit_map": [ + 0.027109479531645775, + 0.018490608781576157, + 0.02433815971016884, + 0.015767857432365417, + 0.02025325782597065, + 0.019601162523031235, + 0.019585780799388885, + 0.01799088902771473, + 0.017855599522590637, + 0.035790350288152695, + 0.02311495877802372, + 0.02936750464141369, + 0.017522981390357018, + 0.013828848488628864, + 0.01800646260380745, + 0.014439396560192108, + 0.015983762219548225, + 0.029092151671648026, + 0.018042907118797302, + 0.01756366342306137, + 0.01989644579589367, + 0.014657673425972462, + 0.012869883328676224, + 0.014333149418234825, + 0.0036199346650391817, + 0.0032684977632015944, + 0.004637859761714935, + 0.0033974817488342524, + 0.003433290868997574, + 0.0021724728867411613, + 0.003729904303327203, + 0.002799500711262226, + 0.013791840523481369, + 0.014315802603960037, + 0.013418762013316154, + 0.01092596910893917, + 0.008344415575265884, + 0.010521207004785538, + 0.01211050245910883, + 0.013480608351528645, + 0.02159721776843071, + 0.005941583309322596, + 0.005027056206017733, + 0.011615492403507233, + 0.006637839134782553, + 0.007636464200913906, + 0.030628664419054985, + 0.008891546167433262, + 0.009239092469215393, + 0.007706987205892801, + 0.00566936144605279, + 0.010128662921488285, + 0.004304827656596899, + 0.003437201026827097, + 0.008275986649096012, + 0.005603214725852013, + 0.00986761599779129, + 0.0069368151016533375, + 0.0062181088142097, + 0.008790644817054272, + 0.0029170603957027197, + 0.002037511207163334, + 0.007327100727707148, + 0.007234552409499884, + 0.006216474808752537, + 0.00515579991042614, + 0.0040865433402359486, + 0.0035386676900088787, + 0.0026750238612294197, + 0.0016204997664317489, + 0.00562177412211895, + 0.0039894613437354565, + 0.06393950432538986, + 0.05863584950566292, + 0.013512487523257732, + 0.005338339600712061, + 0.005521642509847879, + 0.0028382616583257914, + 0.009838291443884373, + 0.0043297926895320415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.009778810665011406 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261252045631409 + ], + "opposite_logprob_tokens": [ + -0.5761252045631409 + ], + "logprob_reward": -0.8261252045631409, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261252045631409, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18013839423656464, + 0.12425631284713745, + 0.12010397762060165, + 0.07504219561815262, + 0.07882893830537796, + 0.0961604118347168, + 0.057193052023649216, + 0.09104519337415695, + 0.07420617341995239, + 0.1030253916978836 + ], + "credit_map": [ + 0.02662496827542782, + 0.024390770122408867, + 0.02522342838346958, + 0.016392961144447327, + 0.029935652390122414, + 0.02129638008773327, + 0.020633455365896225, + 0.015640759840607643, + 0.014787483029067516, + 0.020547959953546524, + 0.01647862046957016, + 0.018577681854367256, + 0.014784760773181915, + 0.012433448806405067, + 0.01502858567982912, + 0.011617769487202168, + 0.011509524658322334, + 0.018106598407030106, + 0.013625390827655792, + 0.016313279047608376, + 0.01156969927251339, + 0.009903861209750175, + 0.01723446138203144, + 0.02184116654098034, + 0.010405010543763638, + 0.00910894200205803, + 0.014904661104083061, + 0.010384585708379745, + 0.007419472560286522, + 0.006750244647264481, + 0.009678862988948822, + 0.006390419322997332, + 0.010107452049851418, + 0.00914385449141264, + 0.017018748447299004, + 0.00880422256886959, + 0.007386044133454561, + 0.008220076560974121, + 0.009645574726164341, + 0.008502962999045849, + 0.018433034420013428, + 0.008215160109102726, + 0.014551841653883457, + 0.014096998609602451, + 0.006130815017968416, + 0.004501220304518938, + 0.019022051244974136, + 0.011209295131266117, + 0.008813568390905857, + 0.006707437802106142, + 0.006136052776128054, + 0.010577593930065632, + 0.00360319041647017, + 0.002779474249109626, + 0.01151701994240284, + 0.00705871544778347, + 0.014947130344808102, + 0.010494337417185307, + 0.010405308566987514, + 0.01578233391046524, + 0.005444603972136974, + 0.004292639903724194, + 0.0182590801268816, + 0.011419759131968021, + 0.01696338690817356, + 0.012815633788704872, + 0.011479390785098076, + 0.007310496177524328, + 0.005782048683613539, + 0.003880661679431796, + 0.009386229328811169, + 0.006588323507457972, + 0.03835170343518257, + 0.03976641967892647, + 0.007920104078948498, + 0.004302015528082848, + 0.003703319700434804, + 0.002387062646448612, + 0.0035964567214250565, + 0.0029983099084347486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.8261252045631409 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, long shot, rendered with the exquisite detail and majestic grandeur of classical Chinese art, captures a pivotal moment of conflict within a vast, dimly lit hall. Massive, dark-lacquered wooden pillars tower over the invisible, high ceiling, while golden light pours through ornate, high windows, permeating the thick atmosphere. At the center of this vast space, a large, heavy Chinese chess board crafted from polished, dark rosewood rests on a low table. The camera zooms in on an ebony chariot, its surface a deep, luminous black, as it violently crashes into an ornately carved ivory commander in breathtaking slow motion. The commander, crafted from aged ivory with a faint yellow patina and adorned with miniature dragon motifs around its base, is instantly launched upward and backward from the board. It tumbles chaotically through a bright beam of light, the crimson ink of the central figure momentarily flashing towards the camera. At the point of impact, a tiny cloud of ivory dust and microscopic debris explodes outward, catching the light like a fleeting, shimmering cloud, while the heavy slabs themselves visibly shimmer, emphasizing the silent, resonant power of the conflict and creating a dramatic spectacle of strategic endgame and controlled violence.", + "seed": 7263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_3/index_2740_seed_7263_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.26170357819646595, + "vlm_logprob": -0.26170357819646595, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.26170357819646595, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012258191592991352 + ], + "opposite_logprob_tokens": [ + -5.387258052825928 + ], + "logprob_reward": -0.012258191592991352, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.012258191592991352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21345530450344086, + 0.07702841609716415, + 0.06019533798098564, + 0.05813983082771301, + 0.07188784331083298, + 0.12720611691474915, + 0.06158167123794556, + 0.023232463747262955, + 0.03046439401805401, + 0.2768086791038513 + ], + "credit_map": [ + 0.030031342059373856, + 0.030024675652384758, + 0.029749799519777298, + 0.03232337906956673, + 0.02768399938941002, + 0.018916014581918716, + 0.021578717976808548, + 0.023147378116846085, + 0.008181767538189888, + 0.011750933714210987, + 0.011529365554451942, + 0.008823045529425144, + 0.009296510368585587, + 0.005212344229221344, + 0.011423549614846706, + 0.010810892097651958, + 0.004476598463952541, + 0.007174805272370577, + 0.011499583721160889, + 0.011982119642198086, + 0.005819261539727449, + 0.0037970668636262417, + 0.008470346219837666, + 0.0069755553267896175, + 0.0056536500342190266, + 0.006270531099289656, + 0.016593504697084427, + 0.015817753970623016, + 0.003781091421842575, + 0.002960096811875701, + 0.0031004631891846657, + 0.0039627389051020145, + 0.016426676884293556, + 0.009117672219872475, + 0.016482913866639137, + 0.00852079875767231, + 0.004909850191324949, + 0.005098622292280197, + 0.004039783496409655, + 0.007291528396308422, + 0.028548704460263252, + 0.02032511867582798, + 0.025071948766708374, + 0.015226684510707855, + 0.01002058107405901, + 0.009774560108780861, + 0.010819731280207634, + 0.00741878105327487, + 0.007600496523082256, + 0.008157680742442608, + 0.008771379478275776, + 0.01181634422391653, + 0.00699071679264307, + 0.005179384723305702, + 0.006849380210042, + 0.006216289009898901, + 0.0033180969767272472, + 0.002048076828941703, + 0.003542287740856409, + 0.003986466210335493, + 0.0024301093071699142, + 0.0015916292322799563, + 0.003451506607234478, + 0.0028642904944717884, + 0.0026311748661100864, + 0.002943748142570257, + 0.004793980158865452, + 0.004343255888670683, + 0.0027548358775675297, + 0.002797777531668544, + 0.005613207817077637, + 0.004586412571370602, + 0.0220772884786129, + 0.01804616115987301, + 0.08256431668996811, + 0.07116513699293137, + 0.016005706042051315, + 0.018297843635082245, + 0.02259962074458599, + 0.02605256997048855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.012258191592991352 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014220098033547401 + ], + "opposite_logprob_tokens": [ + -4.264220237731934 + ], + "logprob_reward": -0.014220098033547401, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014220098033547401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13288632035255432, + 0.04049443081021309, + 0.061416737735271454, + 0.10587495565414429, + 0.07229195535182953, + 0.025435348972678185, + 0.0601884089410305, + 0.01624395325779915, + 0.03794663026928902, + 0.44722121953964233 + ], + "credit_map": [ + 0.0179707370698452, + 0.020351829007267952, + 0.019093671813607216, + 0.016081571578979492, + 0.015449924394488335, + 0.012937277555465698, + 0.01513975765556097, + 0.015861548483371735, + 0.0038494961336255074, + 0.005815914366394281, + 0.00588617566972971, + 0.005161568522453308, + 0.004771190695464611, + 0.0033927091863006353, + 0.0063667879439890385, + 0.00525059225037694, + 0.010065943002700806, + 0.006130825262516737, + 0.010987862944602966, + 0.008752517402172089, + 0.0048164683394134045, + 0.006902438588440418, + 0.007346103433519602, + 0.00641458248719573, + 0.011803800240159035, + 0.005292227491736412, + 0.036311205476522446, + 0.03368180990219116, + 0.004833545070141554, + 0.004521292634308338, + 0.005067657679319382, + 0.004363432060927153, + 0.009432653896510601, + 0.006108507048338652, + 0.01586361601948738, + 0.013890225440263748, + 0.005220419727265835, + 0.005418336484581232, + 0.00907405000180006, + 0.00728415185585618, + 0.0034294924698770046, + 0.0037049446254968643, + 0.0038797964807599783, + 0.004034908022731543, + 0.002591688884422183, + 0.0020162288565188646, + 0.002970595844089985, + 0.0028076942544430494, + 0.010508239269256592, + 0.008179724216461182, + 0.00888098031282425, + 0.012989063747227192, + 0.004507029429078102, + 0.002408180618658662, + 0.007554220035672188, + 0.005160974338650703, + 0.0019501404603943229, + 0.0015352765331044793, + 0.0027724308893084526, + 0.0026090259198099375, + 0.0017219086876139045, + 0.0012474546674638987, + 0.0024826298467814922, + 0.0019250879995524883, + 0.0031652511097490788, + 0.0030557680875062943, + 0.006530182436108589, + 0.0076561858877539635, + 0.0030442585702985525, + 0.0027683305088430643, + 0.007256581448018551, + 0.0044700768776237965, + 0.025484953075647354, + 0.03568132966756821, + 0.13895069062709808, + 0.14050276577472687, + 0.026424765586853027, + 0.018932713195681572, + 0.034364670515060425, + 0.026879362761974335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.014220098033547401 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761592984199524 + ], + "opposite_logprob_tokens": [ + -0.8261592984199524 + ], + "logprob_reward": -0.5761592984199524, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761592984199524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05447172001004219, + 0.01939806528389454, + 0.0351598896086216, + 0.020964810624718666, + 0.04374972730875015, + 0.04914617910981178, + 0.05876319110393524, + 0.010081274434924126, + 0.01929173246026039, + 0.6889734268188477 + ], + "credit_map": [ + 0.008586668409407139, + 0.008997359313070774, + 0.007144180592149496, + 0.005607831757515669, + 0.005142978858202696, + 0.005215127021074295, + 0.00644743861630559, + 0.007330132648348808, + 0.003987078554928303, + 0.0025254006031900644, + 0.0023847080301493406, + 0.0014602275332435966, + 0.0020602778531610966, + 0.0018934202380478382, + 0.0021250112913548946, + 0.002961941761896014, + 0.007418051362037659, + 0.003984611947089434, + 0.004596259444952011, + 0.004646467976272106, + 0.0034146616235375404, + 0.004251484759151936, + 0.0030337583739310503, + 0.003814591094851494, + 0.0028762465808540583, + 0.0031704814173281193, + 0.0030962969176471233, + 0.003368475940078497, + 0.0024290787987411022, + 0.0016630177851766348, + 0.0018517652060836554, + 0.0025094503071159124, + 0.008188373409211636, + 0.003457872662693262, + 0.008324474096298218, + 0.0044378796592354774, + 0.002801412483677268, + 0.003485312918201089, + 0.003609913168475032, + 0.00944448821246624, + 0.003823008621111512, + 0.004644906148314476, + 0.008180311881005764, + 0.010505200363695621, + 0.003569036955013871, + 0.003294974099844694, + 0.008281096816062927, + 0.006847642362117767, + 0.005019483156502247, + 0.004798483103513718, + 0.008195181377232075, + 0.015402765944600105, + 0.0045255147852003574, + 0.0038706872146576643, + 0.009439453482627869, + 0.007511620409786701, + 0.0008364426321350038, + 0.0005999283166602254, + 0.0013008029200136662, + 0.0024373934138566256, + 0.0009669488645158708, + 0.0005848317523486912, + 0.0021580145694315434, + 0.0011969126062467694, + 0.0023941583931446075, + 0.0017816507024690509, + 0.002355601405724883, + 0.003437203587964177, + 0.00196203519590199, + 0.0012289397418498993, + 0.0034247550647705793, + 0.0027073894161731005, + 0.04264472797513008, + 0.04880679398775101, + 0.21016226708889008, + 0.16171501576900482, + 0.05736078694462776, + 0.032847192138433456, + 0.07030937820672989, + 0.06512725353240967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.5761592984199524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535471469163895 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535471469163895, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535471469163895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047379158437252045, + 0.04159122705459595, + 0.038592398166656494, + 0.034226953983306885, + 0.0329696349799633, + 0.0608559176325798, + 0.04142414405941963, + 0.018627872690558434, + 0.0305806752294302, + 0.6537520885467529 + ], + "credit_map": [ + 0.005360713694244623, + 0.006102349609136581, + 0.00651986338198185, + 0.005208906251937151, + 0.005652219522744417, + 0.0035858876071870327, + 0.008412878960371017, + 0.0065363370813429356, + 0.007595394272357225, + 0.0055702761746943, + 0.005450782831758261, + 0.003676246153190732, + 0.004246964119374752, + 0.004147740080952644, + 0.00525742769241333, + 0.005646391771733761, + 0.008966483175754547, + 0.00414982670918107, + 0.004880291875451803, + 0.005403973627835512, + 0.0042340923100709915, + 0.004422011319547892, + 0.0033255955204367638, + 0.0032101196702569723, + 0.00414592120796442, + 0.005558125674724579, + 0.007022755686193705, + 0.005562295205891132, + 0.0025248031597584486, + 0.0026836434844881296, + 0.0026529808528721333, + 0.004076425451785326, + 0.006687860004603863, + 0.004105666186660528, + 0.005463538225740194, + 0.003499253187328577, + 0.002744638826698065, + 0.0027786714490503073, + 0.0038094371557235718, + 0.0038805673830211163, + 0.010725822299718857, + 0.008773703128099442, + 0.008636146783828735, + 0.007671667728573084, + 0.005990162957459688, + 0.005573050584644079, + 0.007103810552507639, + 0.006381551269441843, + 0.005315395537763834, + 0.005713180173188448, + 0.0061950115486979485, + 0.006117967423051596, + 0.004531944170594215, + 0.002582092536613345, + 0.006630811840295792, + 0.00433774059638381, + 0.0019610649906098843, + 0.0021837116219103336, + 0.0031471664551645517, + 0.0039010460022836924, + 0.0016248223837465048, + 0.001159792416729033, + 0.0022002002224326134, + 0.0024500687140971422, + 0.0029862690716981888, + 0.004077211953699589, + 0.0038588643074035645, + 0.005355250556021929, + 0.00286543695256114, + 0.002210772829130292, + 0.005081167444586754, + 0.004145698621869087, + 0.04313437268137932, + 0.05147159472107887, + 0.20557065308094025, + 0.15107767283916473, + 0.04777495935559273, + 0.03279624506831169, + 0.05957775190472603, + 0.062348753213882446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.012535471469163895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933448314666748 + ], + "opposite_logprob_tokens": [ + -0.6933448314666748 + ], + "logprob_reward": -0.6933448314666748, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933448314666748, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1262204647064209, + 0.05843054875731468, + 0.062090449035167694, + 0.08104360103607178, + 0.04492773860692978, + 0.08359348773956299, + 0.046531904488801956, + 0.027346210554242134, + 0.08679335564374924, + 0.3830222189426422 + ], + "credit_map": [ + 0.01299623679369688, + 0.01564493589103222, + 0.0237558763474226, + 0.019017843529582024, + 0.011523101478815079, + 0.012603024952113628, + 0.013891197741031647, + 0.01678825169801712, + 0.005016072653234005, + 0.008522257208824158, + 0.008203978650271893, + 0.006752281915396452, + 0.007756238803267479, + 0.0034916771110147238, + 0.009041891433298588, + 0.00964614562690258, + 0.007411752827465534, + 0.007874988950788975, + 0.011145556345582008, + 0.010149581357836723, + 0.006057101767510176, + 0.003275342984125018, + 0.008682659827172756, + 0.007493461947888136, + 0.009160565212368965, + 0.005024893209338188, + 0.02712935581803322, + 0.02347511798143387, + 0.004126448184251785, + 0.0038189056795090437, + 0.0048263357020914555, + 0.003481972496956587, + 0.007179513573646545, + 0.005998884793370962, + 0.00637792469933629, + 0.006320336367934942, + 0.004615484271198511, + 0.0030062892474234104, + 0.00578960170969367, + 0.005639704875648022, + 0.0169778261333704, + 0.012113058008253574, + 0.012913242913782597, + 0.011711166240274906, + 0.00836462527513504, + 0.006207566242665052, + 0.007307338062673807, + 0.007998664863407612, + 0.004976863507181406, + 0.005128877237439156, + 0.007306488696485758, + 0.010234418325126171, + 0.0035255765542387962, + 0.003124544396996498, + 0.006525884382426739, + 0.005709253251552582, + 0.0033820841927081347, + 0.0020423573441803455, + 0.0036297820042818785, + 0.005563376005738974, + 0.0022525943350046873, + 0.0016305751632899046, + 0.005498824641108513, + 0.003346614073961973, + 0.007404778618365526, + 0.008647180162370205, + 0.012441128492355347, + 0.0109780328348279, + 0.0099216653034091, + 0.00702367490157485, + 0.019857000559568405, + 0.010519896633923054, + 0.03480381891131401, + 0.058922529220581055, + 0.07475189864635468, + 0.05777673423290253, + 0.04772096499800682, + 0.024099010974168777, + 0.047124847769737244, + 0.037822384387254715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.6933448314666748 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, long shot, rendered with the exquisite detail and majestic grandeur of classical Chinese art, captures a pivotal moment of conflict within a vast, dimly lit hall. Massive, dark-lacquered wooden pillars tower over the invisible, high ceiling, while golden light pours through ornate, high windows, permeating the thick atmosphere. At the center of this vast space, a large, heavy Chinese chess board crafted from polished, dark rosewood rests on a low table. The camera zooms in on an ebony chariot, its surface a deep, luminous black, as it violently crashes into an ornately carved ivory commander in breathtaking slow motion. The commander, crafted from aged ivory with a faint yellow patina and adorned with miniature dragon motifs around its base, is instantly launched upward and backward from the board. It tumbles chaotically through a bright beam of light, the crimson ink of the central figure momentarily flashing towards the camera. At the point of impact, a tiny cloud of ivory dust and microscopic debris explodes outward, catching the light like a fleeting, shimmering cloud, while the heavy slabs themselves visibly shimmer, emphasizing the silent, resonant power of the conflict and creating a dramatic spectacle of strategic endgame and controlled violence.", + "seed": 7264, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_3/index_2740_seed_7264_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.29176784213632345, + "vlm_logprob": -0.29176784213632345, + "vlm_yes_no_margin": 2.300000011920929, + "vlm_token_credit": -0.29176784213632345, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.042284127324819565 + ], + "opposite_logprob_tokens": [ + -3.4172840118408203 + ], + "logprob_reward": -0.042284127324819565, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.042284127324819565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06704159826040268, + 0.03209824860095978, + 0.04189939796924591, + 0.059698522090911865, + 0.054792676120996475, + 0.10626205801963806, + 0.19349369406700134, + 0.03728574886918068, + 0.08302178978919983, + 0.324406236410141 + ], + "credit_map": [ + 0.008046956732869148, + 0.010752387344837189, + 0.010631293058395386, + 0.008012219332158566, + 0.007987539283931255, + 0.00747184082865715, + 0.006623486522585154, + 0.007515866309404373, + 0.0043135350570082664, + 0.004545856732875109, + 0.004529651254415512, + 0.003577735973522067, + 0.0026656179688870907, + 0.00505430530756712, + 0.002978363772854209, + 0.00443318160250783, + 0.007403768133372068, + 0.007511593867093325, + 0.007628832012414932, + 0.0058380113914608955, + 0.003651499282568693, + 0.003376531647518277, + 0.004154043737798929, + 0.002335118828341365, + 0.006010835058987141, + 0.004099699202924967, + 0.004502071999013424, + 0.006534873973578215, + 0.012622338719666004, + 0.009411381557583809, + 0.009759505279362202, + 0.006757815834134817, + 0.00464541744440794, + 0.005392592400312424, + 0.009291231632232666, + 0.007881511934101582, + 0.00363126490265131, + 0.00791946705430746, + 0.006907889153808355, + 0.009123302064836025, + 0.012258254922926426, + 0.013609089888632298, + 0.018995292484760284, + 0.015663068741559982, + 0.01192962285131216, + 0.012002840638160706, + 0.011732974089682102, + 0.01007090788334608, + 0.012740050442516804, + 0.015207028947770596, + 0.009168380871415138, + 0.012744155712425709, + 0.014046236872673035, + 0.0509546622633934, + 0.06462447345256805, + 0.014008702710270882, + 0.0036819339729845524, + 0.003771417774260044, + 0.00614531897008419, + 0.008032522164285183, + 0.0028590254951268435, + 0.005244702100753784, + 0.003645289456471801, + 0.0039055366069078445, + 0.005863763391971588, + 0.022079596295952797, + 0.006639221683144569, + 0.009936953894793987, + 0.004701387602835894, + 0.006527565885335207, + 0.008077998645603657, + 0.019195301458239555, + 0.024324936792254448, + 0.04493649676442146, + 0.050409477204084396, + 0.0917673259973526, + 0.01583370752632618, + 0.020727651193737984, + 0.03341512382030487, + 0.0429915189743042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.042284127324819565 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813103586435318 + ], + "opposite_logprob_tokens": [ + -3.28813099861145 + ], + "logprob_reward": -0.03813103586435318, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813103586435318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06904336810112, + 0.052058570086956024, + 0.0878196433186531, + 0.15258023142814636, + 0.032743293792009354, + 0.0563189797103405, + 0.1434832364320755, + 0.02281377837061882, + 0.12727537751197815, + 0.2558634877204895 + ], + "credit_map": [ + 0.007202381268143654, + 0.013009707443416119, + 0.00897150207310915, + 0.009168372489511967, + 0.008067828603088856, + 0.007434817496687174, + 0.008265442214906216, + 0.006923323962837458, + 0.005492830183357, + 0.005365912336856127, + 0.004981525707989931, + 0.006363728083670139, + 0.0067805759608745575, + 0.007258871104568243, + 0.009795124642550945, + 0.006020004861056805, + 0.007193834986537695, + 0.006856285501271486, + 0.008925437927246094, + 0.008907408453524113, + 0.013315633870661259, + 0.012888041324913502, + 0.022026078775525093, + 0.007706927135586739, + 0.017097892239689827, + 0.01491332147270441, + 0.013510088436305523, + 0.012601803988218307, + 0.030401194468140602, + 0.02286653220653534, + 0.02340591512620449, + 0.017783507704734802, + 0.004702454432845116, + 0.004387467168271542, + 0.0026823258958756924, + 0.004119049292057753, + 0.004059070721268654, + 0.005006188061088324, + 0.0034800053108483553, + 0.0043067364022135735, + 0.005964651238173246, + 0.00765811325982213, + 0.005484247580170631, + 0.007286633364856243, + 0.00547461723908782, + 0.006959389895200729, + 0.010041073895990849, + 0.007450256496667862, + 0.010240293107926846, + 0.009921004995703697, + 0.011113542132079601, + 0.013120798394083977, + 0.010456216521561146, + 0.031727615743875504, + 0.0455990806221962, + 0.011304691433906555, + 0.0018672678852453828, + 0.002045932225883007, + 0.003166762413457036, + 0.004650802351534367, + 0.001872716355137527, + 0.00422302819788456, + 0.0029927094001322985, + 0.0019945597741752863, + 0.008638043887913227, + 0.038148172199726105, + 0.009865602478384972, + 0.017225852236151695, + 0.007748588919639587, + 0.007475878112018108, + 0.011339478194713593, + 0.02683377079665661, + 0.017973896116018295, + 0.03692734241485596, + 0.03699108958244324, + 0.0767383724451065, + 0.01215262059122324, + 0.01558970008045435, + 0.027130037546157837, + 0.03236043080687523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.03813103586435318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31346529722213745 + ], + "opposite_logprob_tokens": [ + -1.3134653568267822 + ], + "logprob_reward": -0.31346529722213745, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31346529722213745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16538910567760468, + 0.09645574539899826, + 0.22061002254486084, + 0.17887331545352936, + 0.060597535222768784, + 0.03617885708808899, + 0.03824330493807793, + 0.05702918395400047, + 0.059612832963466644, + 0.08701011538505554 + ], + "credit_map": [ + 0.0206440482288599, + 0.03250984475016594, + 0.021472817286849022, + 0.018281981348991394, + 0.022032152861356735, + 0.019795924425125122, + 0.016464293003082275, + 0.01418804656714201, + 0.011805813759565353, + 0.011333620175719261, + 0.014217568561434746, + 0.00981956534087658, + 0.009494918398559093, + 0.012635068036615849, + 0.013162704184651375, + 0.013986482284963131, + 0.018520459532737732, + 0.02232169732451439, + 0.025388995185494423, + 0.020032666623592377, + 0.027897436171770096, + 0.03718225285410881, + 0.04728112369775772, + 0.02198537439107895, + 0.01288509275764227, + 0.013381986878812313, + 0.018561312928795815, + 0.020998725667595863, + 0.027503253892064095, + 0.02806594967842102, + 0.029706481844186783, + 0.027770504355430603, + 0.005433873739093542, + 0.004913798067718744, + 0.005819372832775116, + 0.00938372127711773, + 0.00479008536785841, + 0.008834978565573692, + 0.009046098217368126, + 0.0123756043612957, + 0.003422918263822794, + 0.0025342178996652365, + 0.007338536437600851, + 0.006028882693499327, + 0.0026686382479965687, + 0.004216392058879137, + 0.005556343588978052, + 0.004412928130477667, + 0.002786366967484355, + 0.003980069421231747, + 0.002554947044700384, + 0.005466145928949118, + 0.0019946692045778036, + 0.0075614675879478455, + 0.009895085357129574, + 0.0040045552887022495, + 0.0055829002521932125, + 0.005320959724485874, + 0.006482752971351147, + 0.00788154173642397, + 0.004161149263381958, + 0.014379232190549374, + 0.006819339469075203, + 0.006401308812201023, + 0.0044154394418001175, + 0.011929740197956562, + 0.006126027088612318, + 0.006811486091464758, + 0.0034303783904761076, + 0.003639806993305683, + 0.008708029985427856, + 0.0145519208163023, + 0.00748021574690938, + 0.00916050560772419, + 0.011537176556885242, + 0.02641133777797222, + 0.004568277858197689, + 0.005614104215055704, + 0.010670039802789688, + 0.011568459682166576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.31346529722213745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076628230512142 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076628230512142, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076628230512142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09735866636037827, + 0.07509011775255203, + 0.11997701227664948, + 0.07051529735326767, + 0.06615401804447174, + 0.05348854884505272, + 0.25449907779693604, + 0.034489527344703674, + 0.07726752758026123, + 0.15116021037101746 + ], + "credit_map": [ + 0.010851062834262848, + 0.015265284106135368, + 0.014236293733119965, + 0.01323047187179327, + 0.014988498762249947, + 0.00882603507488966, + 0.009480700828135014, + 0.010480327531695366, + 0.007802775129675865, + 0.008189943619072437, + 0.011757534928619862, + 0.007803335785865784, + 0.009086769074201584, + 0.010957594029605389, + 0.007838503457605839, + 0.011653662659227848, + 0.015999630093574524, + 0.011508065275847912, + 0.014970663003623486, + 0.01452710386365652, + 0.019114406779408455, + 0.011625470593571663, + 0.02375747077167034, + 0.008474202826619148, + 0.005765322130173445, + 0.006490321829915047, + 0.008519649505615234, + 0.009566251188516617, + 0.011110294610261917, + 0.011963815428316593, + 0.008046312257647514, + 0.009053340181708336, + 0.007307533174753189, + 0.005431759636849165, + 0.006793841253966093, + 0.009769863449037075, + 0.0056123286485672, + 0.009389749728143215, + 0.010318191722035408, + 0.011530756019055843, + 0.007174033205956221, + 0.006373811978846788, + 0.01120280846953392, + 0.007055368274450302, + 0.004877457860857248, + 0.0063620968721807, + 0.005653688684105873, + 0.004789290018379688, + 0.0170628372579813, + 0.02408170886337757, + 0.017473584040999413, + 0.015551243908703327, + 0.018322566524147987, + 0.06763368844985962, + 0.07997022569179535, + 0.014403261244297028, + 0.0033792401663959026, + 0.003104983363300562, + 0.0061574033461511135, + 0.005615487229079008, + 0.0027656315360218287, + 0.004448278341442347, + 0.00421501137316227, + 0.004803495015949011, + 0.007666701916605234, + 0.013042186386883259, + 0.010181710124015808, + 0.013593289069831371, + 0.004883993417024612, + 0.006695115007460117, + 0.0105137899518013, + 0.010690750554203987, + 0.011318929493427277, + 0.0208699069917202, + 0.021593671292066574, + 0.042809776961803436, + 0.00821917038410902, + 0.010354837402701378, + 0.016347447410225868, + 0.019646475091576576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.011076628230512142 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.053882122039795 + ], + "opposite_logprob_tokens": [ + -0.4288821816444397 + ], + "logprob_reward": -1.053882122039795, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.053882122039795, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10058904439210892, + 0.08881276845932007, + 0.08006717264652252, + 0.15270276367664337, + 0.05547845736145973, + 0.05308088660240173, + 0.21066933870315552, + 0.05844791978597641, + 0.12591633200645447, + 0.07423535734415054 + ], + "credit_map": [ + 0.007301980163902044, + 0.01745239831507206, + 0.010279722511768341, + 0.012212547473609447, + 0.01528996229171753, + 0.012689048424363136, + 0.015974093228578568, + 0.009389279410243034, + 0.014306792989373207, + 0.011739392764866352, + 0.011843669228255749, + 0.008805889636278152, + 0.010953033342957497, + 0.009945034980773926, + 0.012532713823020458, + 0.00868623610585928, + 0.009851454757153988, + 0.00938391126692295, + 0.010274043306708336, + 0.009452723897993565, + 0.010962340049445629, + 0.009329920634627342, + 0.0153525834903121, + 0.005460183136165142, + 0.01578984409570694, + 0.020252810791134834, + 0.016617916524410248, + 0.02135380543768406, + 0.016326485201716423, + 0.022953800857067108, + 0.017582371830940247, + 0.021825706586241722, + 0.004563537426292896, + 0.005479397717863321, + 0.007001014892011881, + 0.011144708842039108, + 0.004140018951147795, + 0.008234952576458454, + 0.006660643499344587, + 0.008254177868366241, + 0.004351435694843531, + 0.006306528113782406, + 0.0063741267658770084, + 0.007723437622189522, + 0.004086533095687628, + 0.007211876567453146, + 0.00824724044650793, + 0.008779700845479965, + 0.01722840778529644, + 0.017446154728531837, + 0.01883205771446228, + 0.019220814108848572, + 0.013114889152348042, + 0.04762815311551094, + 0.05946888402104378, + 0.017729956656694412, + 0.005357720423489809, + 0.005486133508384228, + 0.008005551062524319, + 0.011764362454414368, + 0.004269097466021776, + 0.009201370179653168, + 0.008505153469741344, + 0.0058585284277796745, + 0.008670003153383732, + 0.03417683020234108, + 0.007494781166315079, + 0.011415738612413406, + 0.009975728578865528, + 0.00850172620266676, + 0.015005453489720821, + 0.03067605197429657, + 0.006868497934192419, + 0.004817289765924215, + 0.008595426566898823, + 0.024162201210856438, + 0.0037480557803064585, + 0.00578971765935421, + 0.011490888893604279, + 0.008763274177908897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -1.053882122039795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, long shot, rendered with the exquisite detail and majestic grandeur of classical Chinese art, captures a pivotal moment of conflict within a vast, dimly lit hall. Massive, dark-lacquered wooden pillars tower over the invisible, high ceiling, while golden light pours through ornate, high windows, permeating the thick atmosphere. At the center of this vast space, a large, heavy Chinese chess board crafted from polished, dark rosewood rests on a low table. The camera zooms in on an ebony chariot, its surface a deep, luminous black, as it violently crashes into an ornately carved ivory commander in breathtaking slow motion. The commander, crafted from aged ivory with a faint yellow patina and adorned with miniature dragon motifs around its base, is instantly launched upward and backward from the board. It tumbles chaotically through a bright beam of light, the crimson ink of the central figure momentarily flashing towards the camera. At the point of impact, a tiny cloud of ivory dust and microscopic debris explodes outward, catching the light like a fleeting, shimmering cloud, while the heavy slabs themselves visibly shimmer, emphasizing the silent, resonant power of the conflict and creating a dramatic spectacle of strategic endgame and controlled violence.", + "seed": 7265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_3/index_2740_seed_7265_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3253418014384806, + "vlm_logprob": -0.3253418014384806, + "vlm_yes_no_margin": 2.25, + "vlm_token_credit": -0.3253418014384806, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06245355308055878 + ], + "opposite_logprob_tokens": [ + -2.937453508377075 + ], + "logprob_reward": -0.06245355308055878, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06245355308055878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27630090713500977, + 0.06564269214868546, + 0.0362115278840065, + 0.08627192676067352, + 0.02790123224258423, + 0.05693422257900238, + 0.020959770306944847, + 0.07007676362991333, + 0.18587785959243774, + 0.17382311820983887 + ], + "credit_map": [ + 0.018932262435555458, + 0.022932784631848335, + 0.04633306711912155, + 0.031055258587002754, + 0.0393657311797142, + 0.029793664813041687, + 0.04459364712238312, + 0.043294455856084824, + 0.006569189950823784, + 0.007869711145758629, + 0.011420074850320816, + 0.008070569485425949, + 0.005294356495141983, + 0.008015803061425686, + 0.010412229225039482, + 0.007990751415491104, + 0.0033471346832811832, + 0.004955568816512823, + 0.0035963444970548153, + 0.005890895612537861, + 0.005568380001932383, + 0.004006783477962017, + 0.0054472205229103565, + 0.003399194683879614, + 0.009714754298329353, + 0.011628835462033749, + 0.00953635573387146, + 0.011256420984864235, + 0.012797528877854347, + 0.010414759628474712, + 0.010988464578986168, + 0.009934798814356327, + 0.0034431267995387316, + 0.0042818100191652775, + 0.003842686302959919, + 0.0046226996928453445, + 0.0029211260844022036, + 0.002541792346164584, + 0.0033004088327288628, + 0.0029475793708115816, + 0.0037380894646048546, + 0.007936637848615646, + 0.007122071459889412, + 0.009276434779167175, + 0.007629826199263334, + 0.007044605445116758, + 0.005685669369995594, + 0.008500879630446434, + 0.0022252292837947607, + 0.0028246401343494654, + 0.002461915835738182, + 0.002934043761342764, + 0.003501337952911854, + 0.00277713849209249, + 0.002476810710504651, + 0.0017586505273357034, + 0.0076002031564712524, + 0.010220086202025414, + 0.008423887193202972, + 0.016511844471096992, + 0.008576985448598862, + 0.00575117627158761, + 0.007110076490789652, + 0.005882497411221266, + 0.01397943776100874, + 0.015292448922991753, + 0.025112967938184738, + 0.027594976127147675, + 0.01057321298867464, + 0.01131117157638073, + 0.06503952294588089, + 0.016974108293652534, + 0.014063009060919285, + 0.017727229744195938, + 0.010636286810040474, + 0.030613409355282784, + 0.011905765160918236, + 0.018433382734656334, + 0.04330644756555557, + 0.02713756635785103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.06245355308055878 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637394517660141 + ], + "opposite_logprob_tokens": [ + -3.651373863220215 + ], + "logprob_reward": -0.02637394517660141, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637394517660141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13415656983852386, + 0.029407059773802757, + 0.05339767783880234, + 0.4086845815181732, + 0.06453032791614532, + 0.14066985249519348, + 0.029139427468180656, + 0.025268027558922768, + 0.07040395587682724, + 0.044342558830976486 + ], + "credit_map": [ + 0.012416000477969646, + 0.012079717591404915, + 0.023424582555890083, + 0.015591868199408054, + 0.015562851913273335, + 0.014852824620902538, + 0.02338399738073349, + 0.016844717785716057, + 0.0035188356414437294, + 0.0028878699522465467, + 0.00520454254001379, + 0.0037631515879184008, + 0.002100918209180236, + 0.003906826488673687, + 0.0049231224693357944, + 0.00310179078951478, + 0.003903667675331235, + 0.015006042085587978, + 0.006704811006784439, + 0.009278465062379837, + 0.004135278053581715, + 0.0032979059033095837, + 0.004998389165848494, + 0.006073112599551678, + 0.04578135907649994, + 0.05184781178832054, + 0.03848952427506447, + 0.06196525692939758, + 0.06986313313245773, + 0.05250570923089981, + 0.04602871835231781, + 0.04220304638147354, + 0.009679406881332397, + 0.011366385966539383, + 0.010710596106946468, + 0.008813240565359592, + 0.0054146102629601955, + 0.005052140448242426, + 0.006406792905181646, + 0.00708715058863163, + 0.010011921636760235, + 0.02591104432940483, + 0.02056814730167389, + 0.028513923287391663, + 0.015347804874181747, + 0.011399859562516212, + 0.014214718714356422, + 0.01470242254436016, + 0.0024160637985914946, + 0.003710543503984809, + 0.004356925841420889, + 0.004509049002081156, + 0.004528723657131195, + 0.0046419426798820496, + 0.0026916679926216602, + 0.0022845070343464613, + 0.003063536947593093, + 0.003956292290240526, + 0.0030917886178940535, + 0.004656986799091101, + 0.002751546213403344, + 0.002187861129641533, + 0.0030356612987816334, + 0.0025243523996323347, + 0.005643066950142384, + 0.004743116907775402, + 0.008923914283514023, + 0.010690366849303246, + 0.00438781687989831, + 0.004845372401177883, + 0.024080323055386543, + 0.007089970167726278, + 0.0039817956276237965, + 0.006713295821100473, + 0.007024496793746948, + 0.006187353748828173, + 0.003086357843130827, + 0.004552256781607866, + 0.008138947188854218, + 0.004658052232116461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.02637394517660141 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983216285705566 + ], + "opposite_logprob_tokens": [ + -0.5233216285705566 + ], + "logprob_reward": -0.8983216285705566, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983216285705566, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06885228306055069, + 0.03657994791865349, + 0.08901838213205338, + 0.30429568886756897, + 0.09420281648635864, + 0.07753886282444, + 0.04198913648724556, + 0.10898623615503311, + 0.11610124260187149, + 0.062435418367385864 + ], + "credit_map": [ + 0.007933419197797775, + 0.006063616368919611, + 0.013495447114109993, + 0.007756907492876053, + 0.006942371372133493, + 0.008405952714383602, + 0.01095486618578434, + 0.00729970820248127, + 0.002887345850467682, + 0.006162163335829973, + 0.00424216128885746, + 0.005462125409394503, + 0.002893984317779541, + 0.004727703519165516, + 0.005501037463545799, + 0.004703426733613014, + 0.0055499691516160965, + 0.024306710809469223, + 0.01191516499966383, + 0.01747620292007923, + 0.0069412109442055225, + 0.005786641035228968, + 0.009881172329187393, + 0.0071613118052482605, + 0.034779999405145645, + 0.03715423867106438, + 0.027203334495425224, + 0.03760787099599838, + 0.06064155325293541, + 0.041792627424001694, + 0.032768964767456055, + 0.03234708309173584, + 0.016860252246260643, + 0.012755151838064194, + 0.014344675466418266, + 0.013900495134294033, + 0.008071253076195717, + 0.011868189089000225, + 0.008220341056585312, + 0.008182460442185402, + 0.006500452756881714, + 0.013779133558273315, + 0.011982566677033901, + 0.017757883295416832, + 0.006687569431960583, + 0.007944059558212757, + 0.006513642612844706, + 0.006373553536832333, + 0.005233112722635269, + 0.0032043380197137594, + 0.005270239431411028, + 0.003846053034067154, + 0.006164989899843931, + 0.009596310555934906, + 0.00383757590316236, + 0.004836516920477152, + 0.009338374249637127, + 0.01631803996860981, + 0.011782749556005001, + 0.027403760701417923, + 0.010436326265335083, + 0.006397211458534002, + 0.01573069393634796, + 0.01157908421009779, + 0.010760662145912647, + 0.01582646742463112, + 0.01632644049823284, + 0.01512901671230793, + 0.007676782552152872, + 0.010892735794186592, + 0.02847019024193287, + 0.011018939316272736, + 0.0061126151122152805, + 0.007409149315208197, + 0.008478932082653046, + 0.009931294247508049, + 0.005308325402438641, + 0.007412693463265896, + 0.011790983378887177, + 0.00599142350256443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.8983216285705566 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738320458680391 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738320458680391, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738320458680391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09686867892742157, + 0.04159209877252579, + 0.09934566169977188, + 0.34465333819389343, + 0.053492289036512375, + 0.10282812267541885, + 0.029218515381217003, + 0.08063118159770966, + 0.11490903794765472, + 0.03646109625697136 + ], + "credit_map": [ + 0.0062390579842031, + 0.006605776958167553, + 0.0156800989061594, + 0.009081927128136158, + 0.01923397183418274, + 0.00946180708706379, + 0.020206930115818977, + 0.010359104722738266, + 0.003592439228668809, + 0.0065406267531216145, + 0.004205150995403528, + 0.006092838943004608, + 0.0033736624754965305, + 0.004658662248402834, + 0.0066300868056714535, + 0.0064986287616193295, + 0.006335499230772257, + 0.023325461894273758, + 0.009359515272080898, + 0.016631126403808594, + 0.011607029475271702, + 0.008195224218070507, + 0.014294201508164406, + 0.009597609750926495, + 0.039086636155843735, + 0.04257609322667122, + 0.029825085774064064, + 0.04026238992810249, + 0.07387640327215195, + 0.04702886566519737, + 0.03610386699438095, + 0.035893991589546204, + 0.008557094261050224, + 0.006441441364586353, + 0.008004730567336082, + 0.006201066076755524, + 0.005967147648334503, + 0.006965798791497946, + 0.005948780104517937, + 0.005406227894127369, + 0.0075420052744448185, + 0.021471314132213593, + 0.016148140653967857, + 0.02191087044775486, + 0.009054851718246937, + 0.00886390171945095, + 0.00936790183186531, + 0.00846913829445839, + 0.0037675993517041206, + 0.0022634437773376703, + 0.00353323295712471, + 0.002710688626393676, + 0.004442912992089987, + 0.004959203768521547, + 0.003553154645487666, + 0.003988279495388269, + 0.00676944712176919, + 0.010799860581755638, + 0.009965072385966778, + 0.022774912416934967, + 0.00882516149431467, + 0.004633157979696989, + 0.010541466064751148, + 0.0063221026211977005, + 0.013536394573748112, + 0.009489895775914192, + 0.014313602820038795, + 0.013523753732442856, + 0.006422627251595259, + 0.00945994257926941, + 0.033846236765384674, + 0.014316584914922714, + 0.0034845394548028708, + 0.004199498798698187, + 0.00542638311162591, + 0.005764680448919535, + 0.003341062692925334, + 0.0047546555288136005, + 0.0052529312670230865, + 0.004237346351146698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.006738320458680391 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, dimly lit hall with massive dark-lacquered wooden pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large, heavy Chinese chess board made of polished dark rosewood visible on a low table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does an ebony chariot crash violently into an ornately carved ivory commander?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ivory commander launched upward and backward, tumbling through a bright beam of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera zoom in on the collision between the chariot and the commander?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328215599060059 + ], + "opposite_logprob_tokens": [ + -0.7578215599060059 + ], + "logprob_reward": -0.6328215599060059, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328215599060059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15634697675704956, + 0.06745216250419617, + 0.07129005342721939, + 0.29297205805778503, + 0.0529303215444088, + 0.02254120260477066, + 0.09019792824983597, + 0.07675368338823318, + 0.13881945610046387, + 0.030696114525198936 + ], + "credit_map": [ + 0.011180015280842781, + 0.011546286754310131, + 0.023968981578946114, + 0.01829465478658676, + 0.02508482150733471, + 0.015355046838521957, + 0.03212134167551994, + 0.01879584789276123, + 0.005694489926099777, + 0.010153551585972309, + 0.00751179875805974, + 0.01046612299978733, + 0.005977082997560501, + 0.007714459672570229, + 0.011212214827537537, + 0.008722444996237755, + 0.0037060731556266546, + 0.011778288520872593, + 0.0126423928886652, + 0.014491314068436623, + 0.0048302519135177135, + 0.006008894182741642, + 0.0106096426025033, + 0.007223198190331459, + 0.028400149196386337, + 0.03207869455218315, + 0.027393506839871407, + 0.07333102077245712, + 0.034769557416439056, + 0.03459063917398453, + 0.025960849598050117, + 0.03644764795899391, + 0.006574553437530994, + 0.004954208619892597, + 0.007102775853127241, + 0.007812028750777245, + 0.004993734881281853, + 0.007664671633392572, + 0.007878977805376053, + 0.005949373356997967, + 0.001967610791325569, + 0.0034805908799171448, + 0.002435529138892889, + 0.0025656314101070166, + 0.0031217762734740973, + 0.0039897789247334, + 0.0021860080305486917, + 0.002794278087094426, + 0.012085748836398125, + 0.01328195072710514, + 0.012894154526293278, + 0.0145866172388196, + 0.010104678571224213, + 0.007945637218654156, + 0.011899890378117561, + 0.0073992530815303326, + 0.007187753450125456, + 0.009348130784928799, + 0.009106027893722057, + 0.019369585439562798, + 0.009090134873986244, + 0.004592081997543573, + 0.011285039596259594, + 0.0067749242298305035, + 0.01306869089603424, + 0.014459531754255295, + 0.022161660715937614, + 0.020945727825164795, + 0.00910643395036459, + 0.01280019711703062, + 0.028371121734380722, + 0.017906082794070244, + 0.003631512401625514, + 0.0027069959323853254, + 0.0043151043355464935, + 0.003998754546046257, + 0.0035726011265069246, + 0.004571026656776667, + 0.004413010086864233, + 0.0034871106036007404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.6328215599060059 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_4/index_tensor([6490, 6490])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_4/index_tensor([6490, 6490])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5d3f7729ef6d7c366dfc6f4e249ea6ba64c38b2c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_4/index_tensor([6490, 6490])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Static, macro time-lapse footage immerses the viewer in the accelerated life cycle of a strawberry plant, set against a backdrop of rich, dark, fertile soil. The sequence begins with the plant in full bloom, its delicate, pristine white, five-petaled blossom radiating from a yellow center, surrounded by lush, jagged-edged green leaves. As time accelerates, the camera remains stationary, capturing the graceful withering and shedding of the white petals, revealing the slow, deliberate swelling of the flower's base into a small, firm, pale green core. Enclosed in a green, star-shaped cup, this nascent fruit steadily and noticeably expands, its surface texture evolving from smooth to a defined pattern of tiny seeds. The remnants of the faded flower linger for a moment before disappearing. The most dramatic stage unfolds, as the green surface begins a magical, seamless transformation, first blushing to pale white, then to a soft pink, and finally deepening to a rich, vibrant crimson. Bathed in a dynamic yet gentle cycle of natural light, the entire process simulates the passage of a day, culminating in a perfectly shaped, plump, glossy red strawberry, its surface dotted with tiny yellow seeds, hanging heavily from its stem, embodying the serene atmosphere of nature’s wonder and the quiet, beautiful process of fruiting.", + "seed": 9751, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_4/index_6490_seed_9751_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14197977626463398, + "vlm_logprob": -0.14197977626463398, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.14197977626463398, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.697357714176178 + ], + "opposite_logprob_tokens": [ + -0.697357714176178 + ], + "logprob_reward": -0.697357714176178, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.697357714176178, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17048560082912445, + 0.08635719120502472, + 0.13923579454421997, + 0.06551999598741531, + 0.05951574817299843, + 0.03583214804530144, + 0.03757971152663231, + 0.08890099823474884, + 0.16153807938098907, + 0.15503475069999695 + ], + "credit_map": [ + 0.012432684190571308, + 0.013635659590363503, + 0.032009899616241455, + 0.010616439394652843, + 0.01733776181936264, + 0.018408922478556633, + 0.03950509428977966, + 0.026539146900177002, + 0.004837160464376211, + 0.014336689375340939, + 0.009981732815504074, + 0.017359435558319092, + 0.009276945143938065, + 0.008434667252004147, + 0.010445148684084415, + 0.011685410514473915, + 0.00931438896805048, + 0.015609310939908028, + 0.018073881044983864, + 0.019329162314534187, + 0.01757081225514412, + 0.01589280366897583, + 0.0198057908564806, + 0.02363963983952999, + 0.005028863903135061, + 0.00837432686239481, + 0.007748558651655912, + 0.005849697161465883, + 0.0064206779934465885, + 0.007612647954374552, + 0.008557308465242386, + 0.015927914530038834, + 0.005461154505610466, + 0.005327168852090836, + 0.005012670997530222, + 0.005337286274880171, + 0.0046729170717298985, + 0.011607609689235687, + 0.009225835092365742, + 0.012871106155216694, + 0.0033522346056997776, + 0.0033956458792090416, + 0.004386590793728828, + 0.0036434349603950977, + 0.0024708344135433435, + 0.005306846462190151, + 0.004573838319629431, + 0.008702724240720272, + 0.005055267363786697, + 0.004703170619904995, + 0.00458596833050251, + 0.0032837814651429653, + 0.002405284671112895, + 0.00528015848249197, + 0.004634744953364134, + 0.007631334010511637, + 0.005295220762491226, + 0.00512325344607234, + 0.01838749088346958, + 0.008437477983534336, + 0.004839640576392412, + 0.011898047290742397, + 0.010502267628908157, + 0.02441760152578354, + 0.009565606713294983, + 0.011505640111863613, + 0.04268316552042961, + 0.014217160642147064, + 0.008600440807640553, + 0.01280621811747551, + 0.01809057965874672, + 0.04406926780939102, + 0.01278409082442522, + 0.009135405533015728, + 0.04856650531291962, + 0.012348374351859093, + 0.010438241995871067, + 0.015385488048195839, + 0.0124850794672966, + 0.03389156982302666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.697357714176178 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076436190865933895 + ], + "opposite_logprob_tokens": [ + -4.882643699645996 + ], + "logprob_reward": -0.0076436190865933895, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076436190865933895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.309848815202713, + 0.07980221509933472, + 0.12912586331367493, + 0.09396456182003021, + 0.05963379144668579, + 0.019017525017261505, + 0.04017561301589012, + 0.07089347392320633, + 0.15045548975467682, + 0.047082602977752686 + ], + "credit_map": [ + 0.021799439564347267, + 0.028356142342090607, + 0.05157278850674629, + 0.04928934946656227, + 0.024872802197933197, + 0.029355725273489952, + 0.07849165052175522, + 0.02611093781888485, + 0.007460255175828934, + 0.012535959482192993, + 0.01002285536378622, + 0.01642637513577938, + 0.004951224196702242, + 0.005433380138128996, + 0.013163994997739792, + 0.009808167815208435, + 0.008099796250462532, + 0.01698325015604496, + 0.01528498250991106, + 0.023093482479453087, + 0.014064940623939037, + 0.014356685802340508, + 0.021215615794062614, + 0.016027111560106277, + 0.008615201339125633, + 0.01411668211221695, + 0.008404236286878586, + 0.011062181554734707, + 0.011329717002809048, + 0.013528072275221348, + 0.011383776552975178, + 0.015524689108133316, + 0.0074434224516153336, + 0.0037679625675082207, + 0.007284455932676792, + 0.004638247191905975, + 0.003640019102022052, + 0.006123811472207308, + 0.008409284986555576, + 0.018326586112380028, + 0.001571298809722066, + 0.002511087339371443, + 0.001728201750665903, + 0.0020737294107675552, + 0.0017818912165239453, + 0.003832862013950944, + 0.0018220823258161545, + 0.003696371801197529, + 0.003353595035150647, + 0.00309372297488153, + 0.0040800911374390125, + 0.0027842000126838684, + 0.004041370004415512, + 0.011799086816608906, + 0.004453497473150492, + 0.006570050027221441, + 0.006912087555974722, + 0.0058794766664505005, + 0.010935206897556782, + 0.006582567002624273, + 0.005637845490127802, + 0.00932617299258709, + 0.008845853619277477, + 0.016774259507656097, + 0.02232753485441208, + 0.013300182297825813, + 0.04155777767300606, + 0.015192271210253239, + 0.007407577708363533, + 0.008935458958148956, + 0.017313718795776367, + 0.024420958012342453, + 0.007261632476001978, + 0.00413533253595233, + 0.011913183145225048, + 0.004854283295571804, + 0.003173066535964608, + 0.0029840704519301653, + 0.005007324740290642, + 0.007753707468509674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0076436190865933895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822346519678831 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822346519678831, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822346519678831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3292810320854187, + 0.1999216377735138, + 0.05186700448393822, + 0.0884266123175621, + 0.08171598613262177, + 0.022525036707520485, + 0.021389473229646683, + 0.03485006093978882, + 0.07194191962480545, + 0.09808129817247391 + ], + "credit_map": [ + 0.026361605152487755, + 0.023093171417713165, + 0.06332428753376007, + 0.024861926212906837, + 0.027272962033748627, + 0.036571286618709564, + 0.10027876496315002, + 0.027517009526491165, + 0.01676896959543228, + 0.02794949896633625, + 0.022988224402070045, + 0.043360836803913116, + 0.016325388103723526, + 0.021725906059145927, + 0.025590088218450546, + 0.02521270327270031, + 0.002919017570093274, + 0.00571819581091404, + 0.006221508141607046, + 0.00715893879532814, + 0.006388864479959011, + 0.006481451913714409, + 0.009268224239349365, + 0.00771079957485199, + 0.006259458605200052, + 0.012820613570511341, + 0.010790552943944931, + 0.007264633662998676, + 0.009434630163013935, + 0.011314019560813904, + 0.009927349165081978, + 0.0206153467297554, + 0.010733184404671192, + 0.005378695670515299, + 0.010355938225984573, + 0.00418039970099926, + 0.00487368181347847, + 0.007570742629468441, + 0.011613144539296627, + 0.02701018936932087, + 0.0020651910454034805, + 0.0023997449316084385, + 0.0017883351538330317, + 0.0022659979294985533, + 0.00217446219176054, + 0.005643629468977451, + 0.002404502360150218, + 0.003783170599490404, + 0.002109827008098364, + 0.0015120691386982799, + 0.0027120269369333982, + 0.0023078874219208956, + 0.0015963565092533827, + 0.004535213578492403, + 0.0026078401133418083, + 0.004008250776678324, + 0.0017476461362093687, + 0.002345716580748558, + 0.003841676516458392, + 0.00437715882435441, + 0.003060462884604931, + 0.007278746459633112, + 0.00606094766408205, + 0.0061377049423754215, + 0.008769797161221504, + 0.01038279663771391, + 0.016546348109841347, + 0.008003421127796173, + 0.004524209536612034, + 0.005357372108846903, + 0.00650913268327713, + 0.01184883899986744, + 0.014109138399362564, + 0.00830546673387289, + 0.02859092690050602, + 0.009832124225795269, + 0.006657117512077093, + 0.004942481406033039, + 0.00959119200706482, + 0.016052836552262306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.002822346519678831 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015149556566029787 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015149556566029787, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015149556566029787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19688193500041962, + 0.1620757281780243, + 0.12872961163520813, + 0.05396433547139168, + 0.06319291889667511, + 0.02324899286031723, + 0.04781224578619003, + 0.09594383090734482, + 0.05930901691317558, + 0.16884136199951172 + ], + "credit_map": [ + 0.015138586051762104, + 0.013047954067587852, + 0.035449378192424774, + 0.018044838681817055, + 0.018743056803941727, + 0.02078784443438053, + 0.055255226790905, + 0.020415067672729492, + 0.00943610817193985, + 0.02355434186756611, + 0.018129508942365646, + 0.0429227352142334, + 0.012687580659985542, + 0.013088171370327473, + 0.022952865809202194, + 0.01930442824959755, + 0.00671817222610116, + 0.01647099107503891, + 0.010972945019602776, + 0.01564543880522251, + 0.021301208063960075, + 0.019709834828972816, + 0.019209088757634163, + 0.018701929599046707, + 0.004286990966647863, + 0.008028454147279263, + 0.0034631690941751003, + 0.0074044507928192616, + 0.0047008865512907505, + 0.008493705652654171, + 0.009564226493239403, + 0.008022450841963291, + 0.005861268378794193, + 0.0059523810632526875, + 0.008967588655650616, + 0.00742550240829587, + 0.005384831223636866, + 0.009273604489862919, + 0.008167651481926441, + 0.012160087004303932, + 0.0026019036304205656, + 0.0022173230536282063, + 0.0030524295289069414, + 0.003306301310658455, + 0.00175988650880754, + 0.0032570171169936657, + 0.002604331821203232, + 0.004449799191206694, + 0.004775157663971186, + 0.003982878755778074, + 0.009122313000261784, + 0.005796872545033693, + 0.0028970658313483, + 0.006379494443535805, + 0.005742922425270081, + 0.009115541353821754, + 0.004140197765082121, + 0.006353072822093964, + 0.012003552168607712, + 0.016738222911953926, + 0.007747870869934559, + 0.009328451007604599, + 0.021770497784018517, + 0.01786196231842041, + 0.005503104999661446, + 0.004105138126760721, + 0.018741877749562263, + 0.005231983959674835, + 0.00275749946013093, + 0.004059491213411093, + 0.004503620322793722, + 0.014406297355890274, + 0.026575429365038872, + 0.016010358929634094, + 0.04186949133872986, + 0.017325859516859055, + 0.011680828407406807, + 0.011011242866516113, + 0.016774063929915428, + 0.02759408764541149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0015149556566029787 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000560245884116739 + ], + "opposite_logprob_tokens": [ + -7.500560283660889 + ], + "logprob_reward": -0.000560245884116739, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.000560245884116739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15551641583442688, + 0.08275537937879562, + 0.09352955222129822, + 0.16606228053569794, + 0.09681151807308197, + 0.03546901047229767, + 0.032389115542173386, + 0.09681544452905655, + 0.13431316614151, + 0.10633805394172668 + ], + "credit_map": [ + 0.016597485169768333, + 0.016381381079554558, + 0.02362530492246151, + 0.023267589509487152, + 0.012392218224704266, + 0.0181526280939579, + 0.023546593263745308, + 0.021553218364715576, + 0.004921511746942997, + 0.011332417838275433, + 0.009846506640315056, + 0.021024515852332115, + 0.007076492067426443, + 0.005957083310931921, + 0.010949749499559402, + 0.011647111736238003, + 0.005562923848628998, + 0.012113613076508045, + 0.011324564926326275, + 0.00825166143476963, + 0.01603115163743496, + 0.014203974045813084, + 0.007953676395118237, + 0.01808798871934414, + 0.01386851817369461, + 0.027941321954131126, + 0.019465429708361626, + 0.01348938513547182, + 0.015079489909112453, + 0.01799923926591873, + 0.019288811832666397, + 0.038930103182792664, + 0.011684742756187916, + 0.005827815271914005, + 0.012748412787914276, + 0.005683433264493942, + 0.005517948418855667, + 0.009876913391053677, + 0.013319489546120167, + 0.03215276449918747, + 0.0029164867009967566, + 0.0024106465280056, + 0.0048069884069263935, + 0.003032778622582555, + 0.0025237146764993668, + 0.004698387812823057, + 0.004878186620771885, + 0.01020182203501463, + 0.0027400983963161707, + 0.002418259158730507, + 0.005909408442676067, + 0.002413768321275711, + 0.002656353637576103, + 0.007198057137429714, + 0.002966005355119705, + 0.006087166722863913, + 0.0051979245617985725, + 0.0059221344999969006, + 0.017343079671263695, + 0.009551746770739555, + 0.006796541158109903, + 0.013393244706094265, + 0.014104110188782215, + 0.024506671354174614, + 0.01564265415072441, + 0.010469069704413414, + 0.041384704411029816, + 0.01294325478374958, + 0.0063718087039887905, + 0.007588547188788652, + 0.013080883771181107, + 0.026832256466150284, + 0.008999100886285305, + 0.007738633081316948, + 0.03223954141139984, + 0.00915010180324316, + 0.007870134897530079, + 0.01036715880036354, + 0.008229908533394337, + 0.02174348011612892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.000560245884116739 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, macro time-lapse footage immerses the viewer in the accelerated life cycle of a strawberry plant, set against a backdrop of rich, dark, fertile soil. The sequence begins with the plant in full bloom, its delicate, pristine white, five-petaled blossom radiating from a yellow center, surrounded by lush, jagged-edged green leaves. As time accelerates, the camera remains stationary, capturing the graceful withering and shedding of the white petals, revealing the slow, deliberate swelling of the flower's base into a small, firm, pale green core. Enclosed in a green, star-shaped cup, this nascent fruit steadily and noticeably expands, its surface texture evolving from smooth to a defined pattern of tiny seeds. The remnants of the faded flower linger for a moment before disappearing. The most dramatic stage unfolds, as the green surface begins a magical, seamless transformation, first blushing to pale white, then to a soft pink, and finally deepening to a rich, vibrant crimson. Bathed in a dynamic yet gentle cycle of natural light, the entire process simulates the passage of a day, culminating in a perfectly shaped, plump, glossy red strawberry, its surface dotted with tiny yellow seeds, hanging heavily from its stem, embodying the serene atmosphere of nature’s wonder and the quiet, beautiful process of fruiting.", + "seed": 9752, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_4/index_6490_seed_9752_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.21799899807665496, + "vlm_logprob": -0.21799899807665496, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.21799899807665496, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0577590465545654 + ], + "opposite_logprob_tokens": [ + -0.43275904655456543 + ], + "logprob_reward": -1.0577590465545654, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0577590465545654, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.43691080808639526, + 0.048452917486429214, + 0.04700494930148125, + 0.04874689131975174, + 0.12638208270072937, + 0.045500747859478, + 0.10676702111959457, + 0.07283806055784225, + 0.053486209362745285, + 0.013910360634326935 + ], + "credit_map": [ + 0.02786705456674099, + 0.1112821027636528, + 0.055208396166563034, + 0.028425177559256554, + 0.038851119577884674, + 0.04448679834604263, + 0.05592725798487663, + 0.07486287504434586, + 0.005454040598124266, + 0.007517403922975063, + 0.0057868510484695435, + 0.006886035669595003, + 0.00503665953874588, + 0.003925355616956949, + 0.00728688295930624, + 0.006559684872627258, + 0.003410483244806528, + 0.004324902314692736, + 0.0031671575270593166, + 0.010321751236915588, + 0.005277521908283234, + 0.005436887964606285, + 0.007689951453357935, + 0.0073762922547757626, + 0.004212561994791031, + 0.0034701647236943245, + 0.005919300019741058, + 0.006059333216398954, + 0.004820475820451975, + 0.008236618712544441, + 0.008770418353378773, + 0.0072580184787511826, + 0.010377979837357998, + 0.011453245766460896, + 0.009477440267801285, + 0.030414778739213943, + 0.010184081271290779, + 0.010130125097930431, + 0.01618935540318489, + 0.028155086562037468, + 0.004487016703933477, + 0.0038494623731821775, + 0.0034006210044026375, + 0.007569160312414169, + 0.004294217564165592, + 0.004658037796616554, + 0.00786211434751749, + 0.009380117990076542, + 0.010050312615931034, + 0.010383802466094494, + 0.011065834201872349, + 0.014238028787076473, + 0.013192291371524334, + 0.011932714842259884, + 0.016873588785529137, + 0.019030438736081123, + 0.00722406106069684, + 0.009230752475559711, + 0.006088085938245058, + 0.009741459041833878, + 0.007643094751983881, + 0.0070749688893556595, + 0.015300177037715912, + 0.010535463690757751, + 0.0059103150852024555, + 0.0044625489972531796, + 0.00832041259855032, + 0.008006386458873749, + 0.004370016977190971, + 0.0066710882820189, + 0.0077066547237336636, + 0.008038786239922047, + 0.002729945583269, + 0.001687870011664927, + 0.0019028227543458343, + 0.001011875458061695, + 0.0013128217542544007, + 0.0016152327880263329, + 0.001164144603535533, + 0.0024856473319232464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -1.0577590465545654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060554549098015 + ], + "opposite_logprob_tokens": [ + -3.8956055641174316 + ], + "logprob_reward": -0.02060554549098015, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060554549098015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7440539598464966, + 0.058258865028619766, + 0.01674397476017475, + 0.007785959169268608, + 0.013307549990713596, + 0.009148973971605301, + 0.0380113385617733, + 0.06383609026670456, + 0.040371596813201904, + 0.008481670171022415 + ], + "credit_map": [ + 0.04682646691799164, + 0.15632642805576324, + 0.07955055683851242, + 0.04838113859295845, + 0.06681492179632187, + 0.09087587893009186, + 0.12816782295703888, + 0.12711073458194733, + 0.007221372332423925, + 0.012480364181101322, + 0.006200734991580248, + 0.006558269262313843, + 0.005388372577726841, + 0.004510908853262663, + 0.007514903321862221, + 0.00838394183665514, + 0.001775892567820847, + 0.001763487234711647, + 0.0017671852838248014, + 0.0024773452896624804, + 0.002158186864107847, + 0.0017838109051808715, + 0.002271989593282342, + 0.0027460779529064894, + 0.0007186523289419711, + 0.0008146183681674302, + 0.000708577164914459, + 0.001164010027423501, + 0.000909029389731586, + 0.0012198074255138636, + 0.0011413702741265297, + 0.0011098943650722504, + 0.0010418029269203544, + 0.001158732920885086, + 0.0008561556460335851, + 0.0035727089270949364, + 0.0010513040469959378, + 0.001149727962911129, + 0.0015682539669796824, + 0.002908863825723529, + 0.0009480653097853065, + 0.0008508869796060026, + 0.0008384514949284494, + 0.000978565076366067, + 0.0009284682455472648, + 0.001190472743473947, + 0.0016555811744183302, + 0.001758482656441629, + 0.0037150755524635315, + 0.004490419290959835, + 0.0037921969778835773, + 0.005422866903245449, + 0.005491777788847685, + 0.004606558941304684, + 0.0045958105474710464, + 0.005896635353565216, + 0.005789657589048147, + 0.007502066902816296, + 0.006082541774958372, + 0.007527008652687073, + 0.0074044507928192616, + 0.00545367831364274, + 0.01449818816035986, + 0.00957849994301796, + 0.003758481703698635, + 0.00358123448677361, + 0.005651270970702171, + 0.005588915199041367, + 0.00433327304199338, + 0.004749107640236616, + 0.0062811607494950294, + 0.006428152788430452, + 0.0017126185121014714, + 0.0009359281975775957, + 0.001123887370340526, + 0.0005367384292185307, + 0.0007584968698211014, + 0.0011771501740440726, + 0.000817066349554807, + 0.0014197840355336666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02060554549098015 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632025051862001 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632025051862001, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632025051862001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24168391525745392, + 0.12987501919269562, + 0.03952404111623764, + 0.05188968405127525, + 0.1619821935892105, + 0.022523457184433937, + 0.14960913360118866, + 0.03684133291244507, + 0.09918030351400375, + 0.0668908953666687 + ], + "credit_map": [ + 0.020795650780200958, + 0.042041972279548645, + 0.02338346466422081, + 0.018650643527507782, + 0.018966306000947952, + 0.027818717062473297, + 0.0492120124399662, + 0.040815163403749466, + 0.011246916837990284, + 0.01289804931730032, + 0.011466162279248238, + 0.030512042343616486, + 0.010088549926877022, + 0.011004004627466202, + 0.02317560464143753, + 0.019483691081404686, + 0.0034536782186478376, + 0.006968055386096239, + 0.002312151715159416, + 0.007679831236600876, + 0.005895677488297224, + 0.0020744656212627888, + 0.004996983800083399, + 0.006143196020275354, + 0.004915961995720863, + 0.005796659272164106, + 0.003463651752099395, + 0.006336314603686333, + 0.009340811520814896, + 0.004267629235982895, + 0.009135379455983639, + 0.00863327644765377, + 0.02188950777053833, + 0.024528633803129196, + 0.013693018816411495, + 0.02008780837059021, + 0.02181275002658367, + 0.012479428201913834, + 0.018025491386651993, + 0.0294655654579401, + 0.0022439879830926657, + 0.0023613462690263987, + 0.0019097256008535624, + 0.002507634926587343, + 0.0030373893678188324, + 0.002657442120835185, + 0.0035780395846813917, + 0.004227891564369202, + 0.013830145820975304, + 0.017039073631167412, + 0.014944981783628464, + 0.021186290308833122, + 0.021468758583068848, + 0.018640395253896713, + 0.018701817840337753, + 0.023797664791345596, + 0.0051836976781487465, + 0.003921966999769211, + 0.003598489798605442, + 0.004364373162388802, + 0.0033244118094444275, + 0.003820887766778469, + 0.007036234252154827, + 0.005591271910816431, + 0.010182245634496212, + 0.008626910857856274, + 0.014550724998116493, + 0.014558200724422932, + 0.00839891005307436, + 0.013013108633458614, + 0.014813600108027458, + 0.015036596916615963, + 0.012834788300096989, + 0.0073491730727255344, + 0.008181828074157238, + 0.0045508756302297115, + 0.006631429307162762, + 0.009647287428379059, + 0.00685182074084878, + 0.010843697004020214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.007632025051862001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820087829604745 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820087829604745, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820087829604745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27145668864250183, + 0.03953058272600174, + 0.038278140127658844, + 0.02736392244696617, + 0.07910366356372833, + 0.029743030667304993, + 0.2012980431318283, + 0.10312046855688095, + 0.11654096096754074, + 0.09356456995010376 + ], + "credit_map": [ + 0.01646016724407673, + 0.05901015177369118, + 0.030728384852409363, + 0.018595334142446518, + 0.025001706555485725, + 0.03273423761129379, + 0.04221497103571892, + 0.046711716800928116, + 0.00349868624471128, + 0.002934438409283757, + 0.0030871774069964886, + 0.008201503194868565, + 0.0032732379622757435, + 0.003993529360741377, + 0.008043131791055202, + 0.006498880684375763, + 0.002405906328931451, + 0.004656728822737932, + 0.002603965112939477, + 0.005500860046595335, + 0.004916854202747345, + 0.004310352262109518, + 0.007213159929960966, + 0.006670314818620682, + 0.001782876905053854, + 0.0017904365668073297, + 0.0020512493792921305, + 0.003937961999326944, + 0.0027414956130087376, + 0.005211967043578625, + 0.005347589496523142, + 0.004500347655266523, + 0.002785002114251256, + 0.005327008198946714, + 0.005048644263297319, + 0.023918168619275093, + 0.007029843050986528, + 0.0056588719598948956, + 0.010564365424215794, + 0.018771760165691376, + 0.0027160579338669777, + 0.003873378736898303, + 0.001861612661741674, + 0.003453102894127369, + 0.004175594542175531, + 0.0034906023647636175, + 0.00497302133589983, + 0.005199661012738943, + 0.019563745707273483, + 0.023951098322868347, + 0.019057603552937508, + 0.0268848929554224, + 0.028580456972122192, + 0.02564164809882641, + 0.024752913042902946, + 0.03286566585302353, + 0.0074994671158492565, + 0.009184541180729866, + 0.009584791027009487, + 0.016419196501374245, + 0.010790767148137093, + 0.008260277099907398, + 0.02403959445655346, + 0.017341841012239456, + 0.012877371162176132, + 0.009817339479923248, + 0.018980570137500763, + 0.013520183973014355, + 0.013206543400883675, + 0.016127848997712135, + 0.013638499192893505, + 0.018372606486082077, + 0.009588058106601238, + 0.010254775173962116, + 0.017492149025201797, + 0.008025364950299263, + 0.007106959354132414, + 0.01608707569539547, + 0.009757026098668575, + 0.015253166668117046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.002820087829604745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178285456262529 + ], + "opposite_logprob_tokens": [ + -6.75117826461792 + ], + "logprob_reward": -0.001178285456262529, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178285456262529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33739495277404785, + 0.04987593740224838, + 0.036533426493406296, + 0.04385799542069435, + 0.12673434615135193, + 0.03527434170246124, + 0.06838396191596985, + 0.09111528098583221, + 0.04152730107307434, + 0.16930241882801056 + ], + "credit_map": [ + 0.02344129979610443, + 0.07278110086917877, + 0.03873715177178383, + 0.02593260630965233, + 0.029460512101650238, + 0.03916138783097267, + 0.0505351796746254, + 0.05734573304653168, + 0.004252323880791664, + 0.005575578659772873, + 0.004849403630942106, + 0.006812708452343941, + 0.0044592092745006084, + 0.005783144850283861, + 0.0105751259252429, + 0.00756844412535429, + 0.0025365047622472048, + 0.003010367276147008, + 0.003492174204438925, + 0.0048749810084700584, + 0.004178510047495365, + 0.006222231313586235, + 0.0067790537141263485, + 0.005439605563879013, + 0.0027363880071789026, + 0.002878685249015689, + 0.003815864445641637, + 0.006204760167747736, + 0.0038284813053905964, + 0.008074803277850151, + 0.008783115074038506, + 0.007535901386290789, + 0.009459328837692738, + 0.01122608594596386, + 0.00968948658555746, + 0.03251992538571358, + 0.011102675460278988, + 0.008860643953084946, + 0.01578013226389885, + 0.02809608168900013, + 0.0035414400044828653, + 0.0028288699686527252, + 0.0025300392881035805, + 0.006160118617117405, + 0.002909485250711441, + 0.004455988761037588, + 0.005585118662565947, + 0.007263285107910633, + 0.005188707262277603, + 0.007049256470054388, + 0.007832581177353859, + 0.011909346096217632, + 0.008139451034367085, + 0.00545310927554965, + 0.010590028949081898, + 0.012221491895616055, + 0.006599046289920807, + 0.008429903537034988, + 0.009316975250840187, + 0.012153539806604385, + 0.011768955737352371, + 0.00864456407725811, + 0.020145105198025703, + 0.014057198539376259, + 0.006661123130470514, + 0.004651269409805536, + 0.006081977393478155, + 0.003426091279834509, + 0.004460925702005625, + 0.004619876854121685, + 0.0043955156579613686, + 0.007230525370687246, + 0.03331228718161583, + 0.02053707093000412, + 0.017642974853515625, + 0.01281477976590395, + 0.01664557494223118, + 0.02351090870797634, + 0.01689671166241169, + 0.027942122891545296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.001178285456262529 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, macro time-lapse footage immerses the viewer in the accelerated life cycle of a strawberry plant, set against a backdrop of rich, dark, fertile soil. The sequence begins with the plant in full bloom, its delicate, pristine white, five-petaled blossom radiating from a yellow center, surrounded by lush, jagged-edged green leaves. As time accelerates, the camera remains stationary, capturing the graceful withering and shedding of the white petals, revealing the slow, deliberate swelling of the flower's base into a small, firm, pale green core. Enclosed in a green, star-shaped cup, this nascent fruit steadily and noticeably expands, its surface texture evolving from smooth to a defined pattern of tiny seeds. The remnants of the faded flower linger for a moment before disappearing. The most dramatic stage unfolds, as the green surface begins a magical, seamless transformation, first blushing to pale white, then to a soft pink, and finally deepening to a rich, vibrant crimson. Bathed in a dynamic yet gentle cycle of natural light, the entire process simulates the passage of a day, culminating in a perfectly shaped, plump, glossy red strawberry, its surface dotted with tiny yellow seeds, hanging heavily from its stem, embodying the serene atmosphere of nature’s wonder and the quiet, beautiful process of fruiting.", + "seed": 9753, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_4/index_6490_seed_9753_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2316154342261143, + "vlm_logprob": -0.2316154342261143, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.2316154342261143, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1480578184127808 + ], + "opposite_logprob_tokens": [ + -0.39805781841278076 + ], + "logprob_reward": -1.1480578184127808, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1480578184127808, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.047779154032468796, + 0.05495642498135567, + 0.024986572563648224, + 0.05368597432971001, + 0.040540385991334915, + 0.04786791652441025, + 0.24278388917446136, + 0.17683830857276917, + 0.300149142742157, + 0.010412268340587616 + ], + "credit_map": [ + 0.004580451641231775, + 0.00750430952757597, + 0.006770784500986338, + 0.005248529836535454, + 0.006102931220084429, + 0.005894232541322708, + 0.0057366397231817245, + 0.005941271316260099, + 0.0038863688241690397, + 0.011834848672151566, + 0.011965317651629448, + 0.005062083248049021, + 0.007653561420738697, + 0.005442931782454252, + 0.004556118045002222, + 0.004555192776024342, + 0.002804211573675275, + 0.004334290511906147, + 0.0033087022602558136, + 0.0024508009664714336, + 0.003474268363788724, + 0.0027175918221473694, + 0.002262197434902191, + 0.0036345089320093393, + 0.007092200685292482, + 0.005388216115534306, + 0.01352020725607872, + 0.007122487761080265, + 0.005707846488803625, + 0.004064571112394333, + 0.006348239257931709, + 0.004442200064659119, + 0.004762287251651287, + 0.004826430231332779, + 0.006581750698387623, + 0.003979155328124762, + 0.003791844705119729, + 0.004034892190247774, + 0.0053423307836055756, + 0.007221687585115433, + 0.004642509389668703, + 0.003816160373389721, + 0.00778261199593544, + 0.00494387885555625, + 0.006966003216803074, + 0.0037352668587118387, + 0.009362689219415188, + 0.006618795916438103, + 0.03023063950240612, + 0.02234385348856449, + 0.05041731521487236, + 0.04081268608570099, + 0.020165234804153442, + 0.02246454916894436, + 0.03196282684803009, + 0.02438676729798317, + 0.026804745197296143, + 0.01808355003595352, + 0.03474855050444603, + 0.018890559673309326, + 0.01969696395099163, + 0.016752494499087334, + 0.021342670544981956, + 0.020518772304058075, + 0.0501086562871933, + 0.028877370059490204, + 0.04972175508737564, + 0.036663979291915894, + 0.03429477661848068, + 0.024468248710036278, + 0.03694497048854828, + 0.039069388061761856, + 0.001096368650905788, + 0.0007968282443471253, + 0.0018177056917920709, + 0.001246039173565805, + 0.0010007974924519658, + 0.0008305820520035923, + 0.0017589759081602097, + 0.00186497054528445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -1.1480578184127808 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002510968828573823 + ], + "opposite_logprob_tokens": [ + -6.002511024475098 + ], + "logprob_reward": -0.002510968828573823, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002510968828573823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05615539848804474, + 0.06217047572135925, + 0.06707946211099625, + 0.05532937869429588, + 0.04194074496626854, + 0.046929892152547836, + 0.22947628796100616, + 0.16422763466835022, + 0.27187544107437134, + 0.004815306980162859 + ], + "credit_map": [ + 0.005751072894781828, + 0.00820714421570301, + 0.00800189096480608, + 0.005338591057807207, + 0.011127525940537453, + 0.007052549626678228, + 0.005396219901740551, + 0.005280401092022657, + 0.004750970751047134, + 0.01067887619137764, + 0.011540272273123264, + 0.006116124801337719, + 0.010598497465252876, + 0.007209166884422302, + 0.006027584429830313, + 0.005248985718935728, + 0.004165966529399157, + 0.005890630651265383, + 0.023634187877178192, + 0.012241648510098457, + 0.004639222752302885, + 0.003410513047128916, + 0.0037682668771594763, + 0.00932903029024601, + 0.006973668001592159, + 0.004931324627250433, + 0.013122977688908577, + 0.006683493498712778, + 0.006642069201916456, + 0.00396651029586792, + 0.008533673360943794, + 0.004475661553442478, + 0.005213558208197355, + 0.005141181871294975, + 0.006851847283542156, + 0.004127926658838987, + 0.004575412720441818, + 0.004066599998623133, + 0.0047878241166472435, + 0.007176391314715147, + 0.004506864119321108, + 0.003711542347446084, + 0.007805749773979187, + 0.005183097440749407, + 0.006229727994650602, + 0.0035686930641531944, + 0.00963236577808857, + 0.006291853729635477, + 0.02886893041431904, + 0.021384991705417633, + 0.0485258512198925, + 0.03794969990849495, + 0.019307978451251984, + 0.020910074934363365, + 0.02999560907483101, + 0.022533144801855087, + 0.025112565606832504, + 0.015977531671524048, + 0.032805316150188446, + 0.01644834503531456, + 0.017715809866786003, + 0.016235608607530594, + 0.019472593441605568, + 0.020459868013858795, + 0.045146554708480835, + 0.02616465836763382, + 0.04490137845277786, + 0.033226728439331055, + 0.03095211461186409, + 0.022360974922776222, + 0.033407971262931824, + 0.03571504354476929, + 0.0006521465838886797, + 0.0004536614869721234, + 0.0007304115570150316, + 0.0005055752117186785, + 0.0005550569039769471, + 0.0004807336663361639, + 0.0006300328532233834, + 0.0008076890953816473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.002510968828573823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097519442439079 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097519442439079, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097519442439079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04262487590312958, + 0.04804319143295288, + 0.06576099246740341, + 0.08310863375663757, + 0.04858829826116562, + 0.07347358763217926, + 0.20746095478534698, + 0.14651358127593994, + 0.2759541869163513, + 0.008471679873764515 + ], + "credit_map": [ + 0.004252827260643244, + 0.005866224877536297, + 0.006139669567346573, + 0.0036592641845345497, + 0.005271265748888254, + 0.006052914075553417, + 0.005755341146141291, + 0.005627366714179516, + 0.004241880029439926, + 0.007543298881500959, + 0.009170684032142162, + 0.004326450638473034, + 0.007764468435198069, + 0.005413210019469261, + 0.004916715435683727, + 0.004666488617658615, + 0.007108990103006363, + 0.0044634174555540085, + 0.022586990147829056, + 0.012578039430081844, + 0.004543231334537268, + 0.0033636046573519707, + 0.004177422262728214, + 0.0069392952136695385, + 0.008669073693454266, + 0.008713520132005215, + 0.01579376496374607, + 0.010306401178240776, + 0.010719597339630127, + 0.006512903142720461, + 0.01628575287759304, + 0.006107617635279894, + 0.004760310053825378, + 0.0061842454597353935, + 0.0077881524339318275, + 0.005024982150644064, + 0.005204427521675825, + 0.00469377962872386, + 0.006831515114754438, + 0.008100884035229683, + 0.006150508765131235, + 0.0059992242604494095, + 0.01337467785924673, + 0.008328093215823174, + 0.007534296251833439, + 0.005461212247610092, + 0.01905079558491707, + 0.007574781309813261, + 0.023302199319005013, + 0.021241595968604088, + 0.041046638041734695, + 0.03947698697447777, + 0.017746536061167717, + 0.018231820315122604, + 0.026320412755012512, + 0.020094754174351692, + 0.021150996908545494, + 0.015118306502699852, + 0.028326762840151787, + 0.015107997693121433, + 0.016521509736776352, + 0.013407174497842789, + 0.018776671960949898, + 0.01810416392982006, + 0.04607854038476944, + 0.026510024443268776, + 0.04543686658143997, + 0.03365519270300865, + 0.03116721846163273, + 0.022956276312470436, + 0.03349969536066055, + 0.03665037825703621, + 0.0015475235413759947, + 0.000906149041838944, + 0.001547518651932478, + 0.0007987581193447113, + 0.000765422941185534, + 0.0010631930781528354, + 0.0006874743849039078, + 0.0011556401150301099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.004097519442439079 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002490753773599863 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.002490753773599863, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002490753773599863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04559558257460594, + 0.08159028738737106, + 0.0889795571565628, + 0.06993409991264343, + 0.061698198318481445, + 0.10016435384750366, + 0.29410409927368164, + 0.10969392210245132, + 0.12713605165481567, + 0.02110384590923786 + ], + "credit_map": [ + 0.003892153035849333, + 0.0058096544817090034, + 0.00700902147218585, + 0.005421221721917391, + 0.005502949468791485, + 0.005226886365562677, + 0.006605205126106739, + 0.006128490436822176, + 0.00555155286565423, + 0.0211928840726614, + 0.025165114551782608, + 0.008145858533680439, + 0.008098401129245758, + 0.0035748672671616077, + 0.003247639397159219, + 0.006613969802856445, + 0.004998048301786184, + 0.00632623303681612, + 0.03436093032360077, + 0.019131435081362724, + 0.004996718838810921, + 0.003086452605202794, + 0.004037552047520876, + 0.01204218715429306, + 0.006161054130643606, + 0.010307995602488518, + 0.008929228410124779, + 0.009597914293408394, + 0.008085539564490318, + 0.006727101281285286, + 0.010740327648818493, + 0.009384939447045326, + 0.0070437220856547356, + 0.006261961534619331, + 0.00954824686050415, + 0.00918340403586626, + 0.006482340861111879, + 0.004662179388105869, + 0.012616273947060108, + 0.0059000723995268345, + 0.009416005574166775, + 0.008310962468385696, + 0.0162492785602808, + 0.012413649819791317, + 0.012670059688389301, + 0.010392136871814728, + 0.01723625510931015, + 0.013476012274622917, + 0.028945855796337128, + 0.03537513315677643, + 0.0497170090675354, + 0.05616443231701851, + 0.03224002942442894, + 0.025813769549131393, + 0.03360380604863167, + 0.03224404901266098, + 0.014201943762600422, + 0.01058852206915617, + 0.0213474053889513, + 0.011188801378011703, + 0.012695160694420338, + 0.010777795687317848, + 0.011551176197826862, + 0.017343102023005486, + 0.01804778352379799, + 0.012948794290423393, + 0.019038379192352295, + 0.015575066208839417, + 0.015806429088115692, + 0.011471794918179512, + 0.016566060483455658, + 0.01768174022436142, + 0.00296827033162117, + 0.0015741698443889618, + 0.003885507583618164, + 0.0014710314571857452, + 0.0019318752456456423, + 0.002321432111784816, + 0.0018516264390200377, + 0.005099933594465256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.002490753773599863 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009201106731779873 + ], + "opposite_logprob_tokens": [ + -7.000920295715332 + ], + "logprob_reward": -0.0009201106731779873, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009201106731779873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03534447029232979, + 0.05333955958485603, + 0.06044187769293785, + 0.0635438859462738, + 0.041011564433574677, + 0.048240769654512405, + 0.23971083760261536, + 0.15766708552837372, + 0.28818821907043457, + 0.012511669658124447 + ], + "credit_map": [ + 0.003251800313591957, + 0.007402351591736078, + 0.0053138937801122665, + 0.0036388160660862923, + 0.004202039912343025, + 0.004451783839613199, + 0.003470376133918762, + 0.003613412147387862, + 0.0035518319346010685, + 0.012927310541272163, + 0.012631061486899853, + 0.004910855088382959, + 0.007083754520863295, + 0.004538544453680515, + 0.003642189782112837, + 0.004054012708365917, + 0.0033370230812579393, + 0.006148426793515682, + 0.021117614582180977, + 0.0105394022539258, + 0.004526311066001654, + 0.002629714785143733, + 0.002712625777348876, + 0.009430763311684132, + 0.006396535784006119, + 0.006379654165357351, + 0.012395475059747696, + 0.007245907559990883, + 0.009484542533755302, + 0.0042770495638251305, + 0.01236753724515438, + 0.004997177980840206, + 0.0048982505686581135, + 0.004731162451207638, + 0.006735996343195438, + 0.004265132825821638, + 0.003959184978157282, + 0.0038421570789068937, + 0.00603829650208354, + 0.006541382055729628, + 0.004895439837127924, + 0.004012421239167452, + 0.007815034128725529, + 0.005221885163336992, + 0.006911973003298044, + 0.003792297327890992, + 0.008641301654279232, + 0.006950416136533022, + 0.029517505317926407, + 0.02277321182191372, + 0.04994305968284607, + 0.040899328887462616, + 0.020504329353570938, + 0.021804621443152428, + 0.031176360324025154, + 0.023092428222298622, + 0.022763676941394806, + 0.01636102795600891, + 0.030850552022457123, + 0.017205694690346718, + 0.017696209251880646, + 0.0148876728489995, + 0.019581245258450508, + 0.018321003764867783, + 0.047686636447906494, + 0.027754152193665504, + 0.047402847558259964, + 0.034876901656389236, + 0.03322885185480118, + 0.023691685870289803, + 0.03583971783518791, + 0.03770744800567627, + 0.0014858510112389922, + 0.001180764171294868, + 0.0020620208233594894, + 0.0013544221874326468, + 0.0011976542882621288, + 0.0013120606308802962, + 0.0020943356212228537, + 0.0018245611572638154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.0009201106731779873 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Static, macro time-lapse footage immerses the viewer in the accelerated life cycle of a strawberry plant, set against a backdrop of rich, dark, fertile soil. The sequence begins with the plant in full bloom, its delicate, pristine white, five-petaled blossom radiating from a yellow center, surrounded by lush, jagged-edged green leaves. As time accelerates, the camera remains stationary, capturing the graceful withering and shedding of the white petals, revealing the slow, deliberate swelling of the flower's base into a small, firm, pale green core. Enclosed in a green, star-shaped cup, this nascent fruit steadily and noticeably expands, its surface texture evolving from smooth to a defined pattern of tiny seeds. The remnants of the faded flower linger for a moment before disappearing. The most dramatic stage unfolds, as the green surface begins a magical, seamless transformation, first blushing to pale white, then to a soft pink, and finally deepening to a rich, vibrant crimson. Bathed in a dynamic yet gentle cycle of natural light, the entire process simulates the passage of a day, culminating in a perfectly shaped, plump, glossy red strawberry, its surface dotted with tiny yellow seeds, hanging heavily from its stem, embodying the serene atmosphere of nature’s wonder and the quiet, beautiful process of fruiting.", + "seed": 9754, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_4/index_6490_seed_9754_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2653878255630843, + "vlm_logprob": -0.2653878255630843, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.2653878255630843, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.321685791015625 + ], + "opposite_logprob_tokens": [ + -0.321685791015625 + ], + "logprob_reward": -1.321685791015625, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.321685791015625, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07271774113178253, + 0.04284892976284027, + 0.038534220308065414, + 0.11153138428926468, + 0.06021809205412865, + 0.05679693818092346, + 0.10359465330839157, + 0.27106934785842896, + 0.11087490618228912, + 0.13181374967098236 + ], + "credit_map": [ + 0.01347061712294817, + 0.005325902719050646, + 0.012325040064752102, + 0.003386524273082614, + 0.005489584058523178, + 0.011127252131700516, + 0.006064894609153271, + 0.015527931973338127, + 0.005372598767280579, + 0.002822604263201356, + 0.007047427352517843, + 0.0038555648643523455, + 0.004423832520842552, + 0.006727579515427351, + 0.0035352909471839666, + 0.009064036421477795, + 0.0059668743051588535, + 0.00499466760084033, + 0.006751484237611294, + 0.0032543898560106754, + 0.004140933509916067, + 0.005337119568139315, + 0.004217137116938829, + 0.0038716127164661884, + 0.01755908504128456, + 0.01579613797366619, + 0.019051894545555115, + 0.014273985289037228, + 0.010832136496901512, + 0.011063865385949612, + 0.011923383921384811, + 0.011030900292098522, + 0.005248416215181351, + 0.00951443612575531, + 0.007926221005618572, + 0.00890139490365982, + 0.00664236256852746, + 0.006828431971371174, + 0.008950114250183105, + 0.006206715479493141, + 0.004963316023349762, + 0.0057984027080237865, + 0.011882658116519451, + 0.014537286013364792, + 0.0035762828774750233, + 0.004636205732822418, + 0.003985776100307703, + 0.007417014334350824, + 0.009531432762742043, + 0.01809733919799328, + 0.021343963220715523, + 0.028170237317681313, + 0.006439831107854843, + 0.008209429681301117, + 0.005595574155449867, + 0.006206853780895472, + 0.0383937731385231, + 0.03387397527694702, + 0.041513457894325256, + 0.03598392382264137, + 0.03320986405014992, + 0.03264111280441284, + 0.024611545726656914, + 0.030841704457998276, + 0.026668759062886238, + 0.00894509069621563, + 0.018716279417276382, + 0.008223996497690678, + 0.01125443633645773, + 0.008775733411312103, + 0.011911225505173206, + 0.016379399225115776, + 0.01149554643779993, + 0.019856875762343407, + 0.013310506008565426, + 0.007969141937792301, + 0.04234267398715019, + 0.00929071195423603, + 0.017303550615906715, + 0.010244755074381828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -1.321685791015625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008412636234425008 + ], + "opposite_logprob_tokens": [ + -7.12584114074707 + ], + "logprob_reward": -0.0008412636234425008, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008412636234425008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11238420009613037, + 0.050981856882572174, + 0.03246345371007919, + 0.08184219896793365, + 0.06447627395391464, + 0.05589374527335167, + 0.129561185836792, + 0.1706988662481308, + 0.16129395365715027, + 0.14040428400039673 + ], + "credit_map": [ + 0.017407014966011047, + 0.009045146405696869, + 0.035862166434526443, + 0.006192684173583984, + 0.007604741957038641, + 0.01178612932562828, + 0.007603663485497236, + 0.016882656142115593, + 0.007179926615208387, + 0.004373130854219198, + 0.007812283933162689, + 0.004494285210967064, + 0.006159077398478985, + 0.007878478616476059, + 0.004349208436906338, + 0.00873546488583088, + 0.00543093541637063, + 0.003762410720810294, + 0.006003814749419689, + 0.0028905721846967936, + 0.0036337662022560835, + 0.004060349892824888, + 0.003290680469945073, + 0.003390926867723465, + 0.011468439362943172, + 0.011821174994111061, + 0.014161249622702599, + 0.010847141966223717, + 0.007319541182368994, + 0.008279354311525822, + 0.009607905521988869, + 0.008337386883795261, + 0.005445463117212057, + 0.010975350625813007, + 0.009051087312400341, + 0.00969674438238144, + 0.007267090026289225, + 0.006721293553709984, + 0.008476774208247662, + 0.0068424781784415245, + 0.004985445644706488, + 0.006141820456832647, + 0.010763530619442463, + 0.013856755569577217, + 0.004715686198323965, + 0.004550657235085964, + 0.004383340012282133, + 0.006496506277471781, + 0.011318917386233807, + 0.020844915881752968, + 0.026454417034983635, + 0.0348503440618515, + 0.009914956986904144, + 0.010243311524391174, + 0.00768084404990077, + 0.008253482170403004, + 0.023548146709799767, + 0.02522369660437107, + 0.027936922386288643, + 0.018863333389163017, + 0.01821010932326317, + 0.02202949859201908, + 0.015446599572896957, + 0.019440557807683945, + 0.03242442384362221, + 0.012222428806126118, + 0.024488942697644234, + 0.012200811877846718, + 0.017590241506695747, + 0.013345412909984589, + 0.024438509717583656, + 0.024583186954259872, + 0.012538313865661621, + 0.019180187955498695, + 0.014363805763423443, + 0.008935895748436451, + 0.04191882535815239, + 0.010960621759295464, + 0.01771201379597187, + 0.014794616959989071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.0008412636234425008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015163839561864734 + ], + "opposite_logprob_tokens": [ + -6.501516342163086 + ], + "logprob_reward": -0.0015163839561864734, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015163839561864734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051705412566661835, + 0.04488123580813408, + 0.03199061006307602, + 0.0674385204911232, + 0.08155188709497452, + 0.057000093162059784, + 0.07990022003650665, + 0.3702086806297302, + 0.08248843252658844, + 0.13283495604991913 + ], + "credit_map": [ + 0.007264666724950075, + 0.004571294412016869, + 0.013850536197423935, + 0.003188067115843296, + 0.004099526908248663, + 0.007294210139662027, + 0.004221958573907614, + 0.007215146441012621, + 0.006067861802875996, + 0.002765774028375745, + 0.007037400733679533, + 0.004026280250400305, + 0.004394283052533865, + 0.007546923588961363, + 0.003981984686106443, + 0.009060720913112164, + 0.004688488319516182, + 0.003640922950580716, + 0.006000868044793606, + 0.0026367451064288616, + 0.003311291802674532, + 0.0059796348214149475, + 0.0026791859418153763, + 0.0030534681864082813, + 0.008341467939317226, + 0.009741289541125298, + 0.01074405387043953, + 0.008271552622318268, + 0.007664019241929054, + 0.007425809279084206, + 0.007873612456023693, + 0.007376709952950478, + 0.009366284124553204, + 0.01789863221347332, + 0.011344309896230698, + 0.012708543799817562, + 0.006639420986175537, + 0.005794092081487179, + 0.008677507750689983, + 0.009123090654611588, + 0.0056297252885997295, + 0.007120837923139334, + 0.007981843315064907, + 0.008297819644212723, + 0.007418409921228886, + 0.006661903113126755, + 0.008594917133450508, + 0.005294627510011196, + 0.009443342685699463, + 0.01430407352745533, + 0.01412902306765318, + 0.014698763377964497, + 0.008219661191105843, + 0.006999376695603132, + 0.005616931244730949, + 0.006489043589681387, + 0.06164708733558655, + 0.05946790426969528, + 0.06110863760113716, + 0.04004240408539772, + 0.037628356367349625, + 0.039515674114227295, + 0.03311128541827202, + 0.037687286734580994, + 0.015984777361154556, + 0.00613287091255188, + 0.012743729166686535, + 0.0056481449864804745, + 0.007763722445815802, + 0.009368066675961018, + 0.01269955188035965, + 0.012147553265094757, + 0.01165058184415102, + 0.019313378259539604, + 0.013976364396512508, + 0.00844596792012453, + 0.03982047736644745, + 0.01091195922344923, + 0.016644692048430443, + 0.012071521952748299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.0015163839561864734 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017162130679935217 + ], + "opposite_logprob_tokens": [ + -6.376716136932373 + ], + "logprob_reward": -0.0017162130679935217, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017162130679935217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04284460097551346, + 0.024871686473488808, + 0.0414605438709259, + 0.09229041635990143, + 0.07893863320350647, + 0.07321450859308243, + 0.1883353739976883, + 0.2498868703842163, + 0.05992019921541214, + 0.14823713898658752 + ], + "credit_map": [ + 0.009738157503306866, + 0.004014329519122839, + 0.004511324223130941, + 0.0026941190008074045, + 0.004336799029260874, + 0.006367835681885481, + 0.0046171462163329124, + 0.006564893759787083, + 0.003623622003942728, + 0.0018706002738326788, + 0.004053217358887196, + 0.00215208251029253, + 0.0030289655551314354, + 0.0038998634554445744, + 0.002150794258341193, + 0.0040925429202616215, + 0.004574712365865707, + 0.005668923258781433, + 0.006830040831118822, + 0.003877609968185425, + 0.005318937823176384, + 0.005970017984509468, + 0.005455926526337862, + 0.0037643760442733765, + 0.013117014430463314, + 0.013048033230006695, + 0.014507537707686424, + 0.011233983561396599, + 0.010653767734766006, + 0.010703414678573608, + 0.010619410313665867, + 0.008407266810536385, + 0.006482662167400122, + 0.012807492166757584, + 0.01110597513616085, + 0.012923604808747768, + 0.008286028169095516, + 0.007969737984240055, + 0.010635873302817345, + 0.008727268315851688, + 0.006907448172569275, + 0.00758656719699502, + 0.01167502161115408, + 0.012447915971279144, + 0.007400488946586847, + 0.009376605972647667, + 0.009379341267049313, + 0.008441127836704254, + 0.021797655150294304, + 0.03082757629454136, + 0.02718314714729786, + 0.022348998114466667, + 0.023678911849856377, + 0.018253834918141365, + 0.02014027163386345, + 0.024105006828904152, + 0.039206236600875854, + 0.03603273257613182, + 0.03997643664479256, + 0.029689082875847816, + 0.026801560074090958, + 0.030709782615303993, + 0.021821411326527596, + 0.0256496649235487, + 0.009203755296766758, + 0.003932185471057892, + 0.008511149324476719, + 0.004847764037549496, + 0.007717938628047705, + 0.007040363736450672, + 0.012495486997067928, + 0.006171557120978832, + 0.013779180124402046, + 0.02116584964096546, + 0.015846308320760727, + 0.0095589654520154, + 0.041352055966854095, + 0.01569347269833088, + 0.019191455096006393, + 0.011649856343865395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.0017162130679935217 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera stationary throughout the entire sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a white, five-petaled blossom with a yellow center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower wither and reveal a pale green core enclosed in a green, star-shaped cup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the fruit surface transform from pale green to pink and finally to a vibrant crimson red?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final strawberry appear plump, glossy, and dotted with tiny yellow seeds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011794761521741748 + ], + "opposite_logprob_tokens": [ + -6.7511796951293945 + ], + "logprob_reward": -0.0011794761521741748, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011794761521741748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0553460456430912, + 0.02468072436749935, + 0.034768231213092804, + 0.08871836960315704, + 0.07223622500896454, + 0.07400155812501907, + 0.08967660367488861, + 0.20539036393165588, + 0.10639220476150513, + 0.24878966808319092 + ], + "credit_map": [ + 0.012576963752508163, + 0.003927178215235472, + 0.01360421720892191, + 0.003765426343306899, + 0.004748330917209387, + 0.006088496185839176, + 0.004448068328201771, + 0.006187359802424908, + 0.0037803028244525194, + 0.0020167329348623753, + 0.004180483054369688, + 0.002145473612472415, + 0.002640203572809696, + 0.003866643412038684, + 0.002061157952994108, + 0.003989726305007935, + 0.00431720819324255, + 0.004781435243785381, + 0.006018489599227905, + 0.003078542882576585, + 0.003870113519951701, + 0.0049359556287527084, + 0.004555552266538143, + 0.0032109336461871862, + 0.012737033888697624, + 0.0126915592700243, + 0.014158421196043491, + 0.012083473615348339, + 0.008821312338113785, + 0.008685309439897537, + 0.010861103422939777, + 0.008680153638124466, + 0.005809626076370478, + 0.011629092507064342, + 0.010097450576722622, + 0.012499547563493252, + 0.006584166083484888, + 0.007095567416399717, + 0.012001384049654007, + 0.0065193879418075085, + 0.006575234234333038, + 0.008225681260228157, + 0.013588682748377323, + 0.016821222379803658, + 0.006595630664378405, + 0.005563706159591675, + 0.008150146342813969, + 0.008481253869831562, + 0.009501182474195957, + 0.009945807047188282, + 0.018410947173833847, + 0.025278601795434952, + 0.0064170584082603455, + 0.0091492785140872, + 0.005379611626267433, + 0.0055941203609108925, + 0.029965855181217194, + 0.024653436616063118, + 0.03125530108809471, + 0.027506567537784576, + 0.024061892181634903, + 0.026232130825519562, + 0.017474306747317314, + 0.024240875616669655, + 0.02360139600932598, + 0.008911647833883762, + 0.01752677746117115, + 0.007772035431116819, + 0.009764914400875568, + 0.00835928600281477, + 0.012490406632423401, + 0.017965739592909813, + 0.02204113081097603, + 0.035587430000305176, + 0.02462593838572502, + 0.01713002286851406, + 0.0766376480460167, + 0.01859230361878872, + 0.033583540469408035, + 0.020591657608747482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.0011794761521741748 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_5/index_tensor([2504, 2504])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_5/index_tensor([2504, 2504])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d7e1026998db8ecae47635d4198af9465785e56b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_5/index_tensor([2504, 2504])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-contrast black-and-white close-up, shot from a level angle, focuses on the face of an elderly man, marked by the marks of time. Sharp afternoon sunlight streams through a blinded window, casting a crisscross pattern of light and shadow across his furrowed forehead and sunken cheeks. Bright streaks intertwine with deep shadows, sculpting his features into a richly textured relief. His sparse, short, silver-white hair gleams glaringly in the strong light, while his dark cotton shirt, frayed at the collar, blends almost completely into the darkness of the background. His chest rises and falls slowly and heavily, and then his cracked lips part slightly, releasing a nearly inaudible sigh, profound and filled with endless fatigue. Then, his eyelids, as if carrying a thousand-pound weight, began to droop irresistibly and extremely slowly. The crow's feet at the corners of his eyes were deeply squeezed together with this movement, and finally completely covered his cloudy eyes, blocking out all the light and stories. The whole picture was filled with a profound sense of powerlessness and a quiet atmosphere of sadness in the extreme contrast of light and shadow.", + "seed": 18985, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_5/index_2504_seed_18985_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05876660868525505, + "vlm_logprob": -0.05876660868525505, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.05876660868525505, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0676422119140625 + ], + "opposite_logprob_tokens": [ + -3.3176422119140625 + ], + "logprob_reward": -0.0676422119140625, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0676422119140625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09671421349048615, + 0.023857858031988144, + 0.013533524237573147, + 0.014254942536354065, + 0.6684600114822388, + 0.045147620141506195, + 0.008695503696799278, + 0.01396195124834776, + 0.02342475950717926, + 0.09194960445165634 + ], + "credit_map": [ + 0.019889812916517258, + 0.006531263701617718, + 0.01908944547176361, + 0.009555802680552006, + 0.006955976132303476, + 0.005617484450340271, + 0.016612866893410683, + 0.01246156170964241, + 0.0029469942674040794, + 0.004080823622643948, + 0.0028491371776908636, + 0.004148699808865786, + 0.002631246345117688, + 0.0020272848196327686, + 0.0029384882654994726, + 0.002235183259472251, + 0.0013847524533048272, + 0.002216228051111102, + 0.0017247127834707499, + 0.0023474241606891155, + 0.0015245702816173434, + 0.0010557669447734952, + 0.0023133428767323494, + 0.0009667265694588423, + 0.0017040058737620711, + 0.0017004563705995679, + 0.0019013026030734181, + 0.0017581927822902799, + 0.0014753381256014109, + 0.0011624045437201858, + 0.0025419394951313734, + 0.0020113030914217234, + 0.20347166061401367, + 0.020128589123487473, + 0.23700003325939178, + 0.037077710032463074, + 0.016117898747324944, + 0.022870928049087524, + 0.034431565552949905, + 0.09736162424087524, + 0.008083966560661793, + 0.00555640272796154, + 0.005957838147878647, + 0.00829224567860365, + 0.004156298469752073, + 0.0022038978058844805, + 0.006711220368742943, + 0.0041857496835291386, + 0.0015519890002906322, + 0.0009345156140625477, + 0.0011986402096226811, + 0.0009600837947800756, + 0.0010456946911290288, + 0.0006406665779650211, + 0.0010379751911386847, + 0.0013259389670565724, + 0.0021905507892370224, + 0.0010298889828845859, + 0.0016916024032980204, + 0.0012739442754536867, + 0.00206335773691535, + 0.0013060157652944326, + 0.0016341347945854068, + 0.0027724560350179672, + 0.0056162732653319836, + 0.0013975496403872967, + 0.004512018524110317, + 0.0013875066069886088, + 0.0018530472880229354, + 0.0025463050696998835, + 0.0016078798798844218, + 0.004504179581999779, + 0.025848444551229477, + 0.0059096477925777435, + 0.02447807788848877, + 0.005373010877519846, + 0.008190249092876911, + 0.0065902420319616795, + 0.009478767402470112, + 0.0060811638832092285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0676422119140625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020612552762031555 + ], + "opposite_logprob_tokens": [ + -3.8956124782562256 + ], + "logprob_reward": -0.020612552762031555, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020612552762031555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1510940045118332, + 0.05586771294474602, + 0.03786814212799072, + 0.026568299159407616, + 0.23978693783283234, + 0.1945151537656784, + 0.04619832709431648, + 0.10256678611040115, + 0.046128880232572556, + 0.09940582513809204 + ], + "credit_map": [ + 0.025928135961294174, + 0.01855899579823017, + 0.022871600463986397, + 0.022159727290272713, + 0.01570410095155239, + 0.009026829153299332, + 0.023216621950268745, + 0.013627984561026096, + 0.008021188899874687, + 0.008048711344599724, + 0.006424862425774336, + 0.006498338188976049, + 0.006106821820139885, + 0.004606922622770071, + 0.007289903704077005, + 0.008870956487953663, + 0.005684659816324711, + 0.0055471244268119335, + 0.004775592125952244, + 0.004549869801849127, + 0.004135775845497847, + 0.003221614519134164, + 0.004337812773883343, + 0.005615689791738987, + 0.004087500739842653, + 0.0035306664649397135, + 0.003485527355223894, + 0.003104021307080984, + 0.0031649479642510414, + 0.003026844235137105, + 0.0032488987781107426, + 0.002919889520853758, + 0.07030616700649261, + 0.00802120752632618, + 0.08196468651294708, + 0.015988139435648918, + 0.00782434456050396, + 0.007712196093052626, + 0.014391588978469372, + 0.03357858210802078, + 0.03219924494624138, + 0.023850535973906517, + 0.02310434728860855, + 0.03510844334959984, + 0.022885456681251526, + 0.012549067847430706, + 0.031719762831926346, + 0.013098281808197498, + 0.007698284462094307, + 0.0060116685926914215, + 0.0060487729497253895, + 0.005141329485923052, + 0.005333529785275459, + 0.004779269453138113, + 0.0073376609943807125, + 0.003847808111459017, + 0.016160709783434868, + 0.013193810358643532, + 0.011975754983723164, + 0.012066847644746304, + 0.01571575738489628, + 0.005166234448552132, + 0.015982143580913544, + 0.01230551116168499, + 0.012198319658637047, + 0.0024124172050505877, + 0.010212735272943974, + 0.0026425933465361595, + 0.0026888460852205753, + 0.004703108221292496, + 0.0032700735609978437, + 0.008000784553587437, + 0.0224546380341053, + 0.008740346878767014, + 0.021960867568850517, + 0.007764057256281376, + 0.012088949792087078, + 0.006958459969609976, + 0.013979953713715076, + 0.00545854214578867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.020612552762031555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029838357120752335 + ], + "opposite_logprob_tokens": [ + -3.5298383235931396 + ], + "logprob_reward": -0.029838357120752335, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029838357120752335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3079775869846344, + 0.06710433214902878, + 0.04051351174712181, + 0.029780900105834007, + 0.248989537358284, + 0.02928769215941429, + 0.030580095946788788, + 0.03161337971687317, + 0.10732132941484451, + 0.10683158785104752 + ], + "credit_map": [ + 0.05247044190764427, + 0.0339202918112278, + 0.05134182423353195, + 0.044356875121593475, + 0.026774149388074875, + 0.018644895404577255, + 0.05713091790676117, + 0.023338211700320244, + 0.010577419772744179, + 0.006340437568724155, + 0.009703435935080051, + 0.004968948662281036, + 0.006370164919644594, + 0.01132066547870636, + 0.009347202256321907, + 0.00847604963928461, + 0.005870860069990158, + 0.0038284657057374716, + 0.00561533635482192, + 0.0029528483282774687, + 0.0036737462505698204, + 0.008048982359468937, + 0.0050611007027328014, + 0.00546216918155551, + 0.0036075362004339695, + 0.002737794304266572, + 0.0036110798828303814, + 0.0025830992963165045, + 0.0028473972342908382, + 0.00477719958871603, + 0.005118393339216709, + 0.004498399794101715, + 0.06804974377155304, + 0.010396945290267467, + 0.07858745753765106, + 0.02019929513335228, + 0.007952847518026829, + 0.009995420463383198, + 0.01729859784245491, + 0.03650923818349838, + 0.004503874108195305, + 0.003560267621651292, + 0.003360504051670432, + 0.003031272441148758, + 0.0036150305531919003, + 0.003252721391618252, + 0.003403186332434416, + 0.004560837056487799, + 0.005855205003172159, + 0.0030412294436246157, + 0.004643693100661039, + 0.0035791362170130014, + 0.0032693222165107727, + 0.0035968138836324215, + 0.003979289438575506, + 0.0026154094375669956, + 0.005330389831215143, + 0.0034749233163893223, + 0.003967625554651022, + 0.0036795029882341623, + 0.0037826814223080873, + 0.0020347624085843563, + 0.004537465050816536, + 0.004806031007319689, + 0.02616293914616108, + 0.0072957659140229225, + 0.02113209292292595, + 0.006885494105517864, + 0.006712906528264284, + 0.01133743580430746, + 0.007885958068072796, + 0.019908741116523743, + 0.027505090460181236, + 0.007133009377866983, + 0.02683096006512642, + 0.007811998017132282, + 0.010446902364492416, + 0.007102330215275288, + 0.011802328750491142, + 0.008198973722755909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.029838357120752335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707367539405823 + ], + "opposite_logprob_tokens": [ + -2.1270737648010254 + ], + "logprob_reward": -0.12707367539405823, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707367539405823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1939951777458191, + 0.0951262041926384, + 0.04674495756626129, + 0.04371641203761101, + 0.12402313947677612, + 0.11061495542526245, + 0.07699862122535706, + 0.060899075120687485, + 0.09031633287668228, + 0.157565176486969 + ], + "credit_map": [ + 0.03151819854974747, + 0.0183931402862072, + 0.02978186309337616, + 0.022269662469625473, + 0.019528193399310112, + 0.013234921731054783, + 0.0325518399477005, + 0.02671734243631363, + 0.014883410185575485, + 0.00990646705031395, + 0.015994424000382423, + 0.013288123533129692, + 0.00944316666573286, + 0.006376216188073158, + 0.014185269363224506, + 0.011049115099012852, + 0.00704072555527091, + 0.005301548633724451, + 0.007436469197273254, + 0.0063611180521547794, + 0.005069939885288477, + 0.004970656242221594, + 0.006228776648640633, + 0.0043357182294130325, + 0.0055906507186591625, + 0.00633862754330039, + 0.005041659809648991, + 0.007225391920655966, + 0.004923963453620672, + 0.004654345568269491, + 0.00547800725325942, + 0.004463762976229191, + 0.032798897475004196, + 0.0058087995275855064, + 0.03806154429912567, + 0.008784997276961803, + 0.005202645901590586, + 0.007495863363146782, + 0.009783742018043995, + 0.0160866416990757, + 0.01651938445866108, + 0.01462949812412262, + 0.012476548552513123, + 0.019996991381049156, + 0.012765772640705109, + 0.008161632344126701, + 0.017192307859659195, + 0.008872817270457745, + 0.012277349829673767, + 0.007948928512632847, + 0.011393249034881592, + 0.009601447731256485, + 0.008186671882867813, + 0.004662493243813515, + 0.01388633530586958, + 0.00904214009642601, + 0.009382843971252441, + 0.0057558827102184296, + 0.00875161588191986, + 0.007699121721088886, + 0.008320026099681854, + 0.003605177626013756, + 0.010913770645856857, + 0.006470628548413515, + 0.029057463631033897, + 0.004711643327027559, + 0.02483174204826355, + 0.006129366811364889, + 0.005843490827828646, + 0.006403004284948111, + 0.005217725411057472, + 0.008121886290609837, + 0.04871527478098869, + 0.010501694865524769, + 0.044952038675546646, + 0.00965123064815998, + 0.013073952868580818, + 0.008066918700933456, + 0.014115209691226482, + 0.008488846011459827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.12707367539405823 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048666246235370636 + ], + "opposite_logprob_tokens": [ + -3.04866623878479 + ], + "logprob_reward": -0.048666246235370636, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048666246235370636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25981205701828003, + 0.059757065027952194, + 0.04072348400950432, + 0.026412280276417732, + 0.3101882338523865, + 0.0371943935751915, + 0.05277910828590393, + 0.05353594943881035, + 0.046824149787425995, + 0.1127733588218689 + ], + "credit_map": [ + 0.04084654897451401, + 0.029863182455301285, + 0.04136214405298233, + 0.031085487455129623, + 0.02610715478658676, + 0.01435483992099762, + 0.04879041016101837, + 0.02740226499736309, + 0.009047562256455421, + 0.007168337237089872, + 0.008693612180650234, + 0.0069476836360991, + 0.005500488448888063, + 0.003808736801147461, + 0.010075831785798073, + 0.008514809422194958, + 0.00523418141528964, + 0.005699048284441233, + 0.004868817050009966, + 0.0057179685682058334, + 0.004934629425406456, + 0.002933523850515485, + 0.0065410383976995945, + 0.004794273525476456, + 0.004398120567202568, + 0.0032369995024055243, + 0.003480852348729968, + 0.00347828958183527, + 0.0030277075711637735, + 0.0028719238471239805, + 0.0033326155971735716, + 0.0025857710279524326, + 0.09422469884157181, + 0.009334598667919636, + 0.10967250168323517, + 0.01588812656700611, + 0.007636436261236668, + 0.01211902592331171, + 0.014939641579985619, + 0.04637317731976509, + 0.006976353470236063, + 0.00458182068541646, + 0.004748705308884382, + 0.0052650016732513905, + 0.004132713656872511, + 0.002283097244799137, + 0.006178473122417927, + 0.0030282263178378344, + 0.008644639514386654, + 0.007642931770533323, + 0.006889629177749157, + 0.006443042773753405, + 0.006345070898532867, + 0.005877504125237465, + 0.007331993896514177, + 0.003604290308430791, + 0.009360821917653084, + 0.006563588511198759, + 0.005859986878931522, + 0.00515733752399683, + 0.006000186782330275, + 0.004553756210952997, + 0.005376131273806095, + 0.010664134286344051, + 0.00984664261341095, + 0.003859828459098935, + 0.007929339073598385, + 0.004201178438961506, + 0.0041277604177594185, + 0.004206794314086437, + 0.004312023986130953, + 0.008340579457581043, + 0.034909624606370926, + 0.007568840403109789, + 0.033022064715623856, + 0.007491035386919975, + 0.008605473674833775, + 0.006917004473507404, + 0.009483668953180313, + 0.004775633569806814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.048666246235370636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-contrast black-and-white close-up, shot from a level angle, focuses on the face of an elderly man, marked by the marks of time. Sharp afternoon sunlight streams through a blinded window, casting a crisscross pattern of light and shadow across his furrowed forehead and sunken cheeks. Bright streaks intertwine with deep shadows, sculpting his features into a richly textured relief. His sparse, short, silver-white hair gleams glaringly in the strong light, while his dark cotton shirt, frayed at the collar, blends almost completely into the darkness of the background. His chest rises and falls slowly and heavily, and then his cracked lips part slightly, releasing a nearly inaudible sigh, profound and filled with endless fatigue. Then, his eyelids, as if carrying a thousand-pound weight, began to droop irresistibly and extremely slowly. The crow's feet at the corners of his eyes were deeply squeezed together with this movement, and finally completely covered his cloudy eyes, blocking out all the light and stories. The whole picture was filled with a profound sense of powerlessness and a quiet atmosphere of sadness in the extreme contrast of light and shadow.", + "seed": 18986, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_5/index_2504_seed_18986_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0729815810918808, + "vlm_logprob": -0.0729815810918808, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.0729815810918808, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04616725817322731 + ], + "opposite_logprob_tokens": [ + -4.421167373657227 + ], + "logprob_reward": -0.04616725817322731, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.04616725817322731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17858527600765228, + 0.11767808347940445, + 0.040123600512742996, + 0.051094017922878265, + 0.05653587728738785, + 0.0752158984541893, + 0.04778808727860451, + 0.046942610293626785, + 0.2845217287540436, + 0.1015147864818573 + ], + "credit_map": [ + 0.028083736076951027, + 0.01125309057533741, + 0.02824992500245571, + 0.015041885897517204, + 0.01572033204138279, + 0.030034568160772324, + 0.026445696130394936, + 0.02375604771077633, + 0.019091760739684105, + 0.005411544349044561, + 0.02076861448585987, + 0.008209573104977608, + 0.02218775451183319, + 0.011079302057623863, + 0.017812363803386688, + 0.013117173686623573, + 0.00685989111661911, + 0.0026191307697445154, + 0.006687183398753405, + 0.0037396298721432686, + 0.005015059839934111, + 0.004694711416959763, + 0.006341986358165741, + 0.004166008904576302, + 0.010495001450181007, + 0.0020639454014599323, + 0.011707997880876064, + 0.004662787076085806, + 0.0058258636854588985, + 0.005462324246764183, + 0.006942341569811106, + 0.003933759871870279, + 0.016738422214984894, + 0.002432455075904727, + 0.015736134722828865, + 0.003758263075724244, + 0.004104064777493477, + 0.003900601528584957, + 0.005059037823230028, + 0.00480689899995923, + 0.007667213212698698, + 0.00406789081171155, + 0.012914172373712063, + 0.004034940619021654, + 0.01309970859438181, + 0.015387894585728645, + 0.008949152193963528, + 0.009094931185245514, + 0.003438754705712199, + 0.0026107800658792257, + 0.006058613769710064, + 0.0028265160508453846, + 0.008711582981050014, + 0.011791973374783993, + 0.009080387651920319, + 0.003269483568146825, + 0.005774087738245726, + 0.0032269787043333054, + 0.007755627855658531, + 0.008362486958503723, + 0.00498479139059782, + 0.007314667105674744, + 0.006362614221870899, + 0.003161355387419462, + 0.07194161415100098, + 0.05272817239165306, + 0.043990179896354675, + 0.021494464948773384, + 0.023623956367373466, + 0.0377470962703228, + 0.021920137107372284, + 0.011076118797063828, + 0.016108829528093338, + 0.006809850689023733, + 0.029005317017436028, + 0.010654548183083534, + 0.009563576430082321, + 0.012707071378827095, + 0.011208991520106792, + 0.005456598475575447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.04616725817322731 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048676010221242905 + ], + "opposite_logprob_tokens": [ + -3.048676013946533 + ], + "logprob_reward": -0.048676010221242905, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048676010221242905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2788725793361664, + 0.11857544630765915, + 0.02802344225347042, + 0.06080343946814537, + 0.23218344151973724, + 0.08499722182750702, + 0.021106930449604988, + 0.06756889820098877, + 0.07827775925397873, + 0.02959078922867775 + ], + "credit_map": [ + 0.04338788241147995, + 0.009869563393294811, + 0.04079195857048035, + 0.011562739498913288, + 0.044482722878456116, + 0.03874364495277405, + 0.05539146810770035, + 0.03464258462190628, + 0.018286436796188354, + 0.004073835909366608, + 0.023421483114361763, + 0.0067595248110592365, + 0.021499445661902428, + 0.01140742190182209, + 0.019773347303271294, + 0.013353952206671238, + 0.0043263486586511135, + 0.0014634852996096015, + 0.0049482546746730804, + 0.0019812677055597305, + 0.003946529235690832, + 0.003221566556021571, + 0.004476346541196108, + 0.003659644164144993, + 0.012378370389342308, + 0.0038889336865395308, + 0.01335914246737957, + 0.005889859050512314, + 0.006260981783270836, + 0.007563717197626829, + 0.007784212473779917, + 0.0036782226525247097, + 0.08501967787742615, + 0.006274642888456583, + 0.08425168693065643, + 0.008442476391792297, + 0.011460930109024048, + 0.012396011501550674, + 0.012942144647240639, + 0.011395872570574284, + 0.007909842766821384, + 0.004904536064714193, + 0.01202174462378025, + 0.0055628614500164986, + 0.013679618947207928, + 0.015032731927931309, + 0.012098930776119232, + 0.013786952011287212, + 0.0028315293602645397, + 0.0019469528924673796, + 0.003238044213503599, + 0.0017446514684706926, + 0.002463635290041566, + 0.004134086426347494, + 0.0030541822779923677, + 0.0016938492190092802, + 0.004416542127728462, + 0.004363826010376215, + 0.010236128233373165, + 0.008692288771271706, + 0.012383327819406986, + 0.009822016581892967, + 0.01211291179060936, + 0.005541857331991196, + 0.016206752508878708, + 0.012762612663209438, + 0.013377382420003414, + 0.006651917938143015, + 0.008393594063818455, + 0.009629372507333755, + 0.007589142769575119, + 0.0036669843830168247, + 0.002953961258754134, + 0.002098384313285351, + 0.006129225715994835, + 0.0029298169538378716, + 0.0051259687170386314, + 0.0036965925246477127, + 0.004700609017163515, + 0.0019562302622944117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.048676010221242905 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020607180893421173 + ], + "opposite_logprob_tokens": [ + -3.8956072330474854 + ], + "logprob_reward": -0.020607180893421173, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020607180893421173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2213309109210968, + 0.05462312698364258, + 0.04305526614189148, + 0.051499173045158386, + 0.04631364345550537, + 0.042192038148641586, + 0.09389137476682663, + 0.07104601711034775, + 0.33883705735206604, + 0.03721148520708084 + ], + "credit_map": [ + 0.04011199250817299, + 0.012896008789539337, + 0.03913207724690437, + 0.027056924998760223, + 0.023644153028726578, + 0.02417091466486454, + 0.034486331045627594, + 0.019832508638501167, + 0.011035734787583351, + 0.004867398180067539, + 0.009209730662405491, + 0.008052986115217209, + 0.005756893195211887, + 0.004709343891590834, + 0.005617058835923672, + 0.00537398224696517, + 0.008803820237517357, + 0.0029501873068511486, + 0.007919011637568474, + 0.005825260654091835, + 0.004854537080973387, + 0.0031090567354112864, + 0.005032420624047518, + 0.004560967907309532, + 0.008870963007211685, + 0.0026837263721972704, + 0.009573116898536682, + 0.004117803648114204, + 0.007959808222949505, + 0.006343777757138014, + 0.008118445053696632, + 0.0038315304554998875, + 0.009301249869167805, + 0.0023528453893959522, + 0.009408288635313511, + 0.004087837878614664, + 0.005453283432871103, + 0.004571301396936178, + 0.006496135611087084, + 0.004642701242119074, + 0.005036494694650173, + 0.002280984539538622, + 0.005588020663708448, + 0.002793699037283659, + 0.005850416608154774, + 0.00802027340978384, + 0.0075613223016262054, + 0.00506082596257329, + 0.011473595164716244, + 0.009080886840820312, + 0.01774946227669716, + 0.011503610759973526, + 0.010963070206344128, + 0.012960868887603283, + 0.013769567012786865, + 0.0063903117552399635, + 0.010491512715816498, + 0.005294215399771929, + 0.02027394436299801, + 0.006928722839802504, + 0.005172536242753267, + 0.011083678342401981, + 0.008210545405745506, + 0.0035908573772758245, + 0.08631961047649384, + 0.061457518488168716, + 0.05562274903059006, + 0.025875814259052277, + 0.02773369289934635, + 0.04456173628568649, + 0.0243617445230484, + 0.012904197908937931, + 0.004022339824587107, + 0.0026929404120892286, + 0.004688689019531012, + 0.0042311870492994785, + 0.006658203434199095, + 0.004748665727674961, + 0.0067872051149606705, + 0.0033822529949247837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.020607180893421173 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039389371871948 + ], + "opposite_logprob_tokens": [ + -1.9103939533233643 + ], + "logprob_reward": -0.16039389371871948, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039389371871948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21170033514499664, + 0.06573968380689621, + 0.06745585799217224, + 0.07849051058292389, + 0.05274507775902748, + 0.13523167371749878, + 0.13171735405921936, + 0.0212563443928957, + 0.1717458814382553, + 0.06391723453998566 + ], + "credit_map": [ + 0.032478030771017075, + 0.008388635702431202, + 0.03170356899499893, + 0.010556190274655819, + 0.03149222582578659, + 0.03412928804755211, + 0.04547737166285515, + 0.017475025728344917, + 0.009063102304935455, + 0.0028463206253945827, + 0.010529175400733948, + 0.00513453409075737, + 0.010927035473287106, + 0.009213310666382313, + 0.011246627196669579, + 0.00677958270534873, + 0.015639852732419968, + 0.002180584706366062, + 0.015978112816810608, + 0.003316151909530163, + 0.007357785012573004, + 0.00894593819975853, + 0.007745489943772554, + 0.0062919436022639275, + 0.014937021769583225, + 0.0024604464415460825, + 0.01728159561753273, + 0.004410445224493742, + 0.010404926724731922, + 0.009769100695848465, + 0.012624609284102917, + 0.006602364592254162, + 0.010062169283628464, + 0.002439554315060377, + 0.011322544887661934, + 0.004239368252456188, + 0.004845272284001112, + 0.005214712116867304, + 0.00803352426737547, + 0.006587933748960495, + 0.011979683302342892, + 0.0053796228021383286, + 0.03112763725221157, + 0.007674883119761944, + 0.012584301643073559, + 0.032441768795251846, + 0.018268804997205734, + 0.015774952247738838, + 0.01650216057896614, + 0.009184326976537704, + 0.02951558120548725, + 0.015035768039524555, + 0.013455025851726532, + 0.017489155754446983, + 0.018981842324137688, + 0.011553476564586163, + 0.0024996281135827303, + 0.001655958709307015, + 0.0023593695368617773, + 0.0020031705498695374, + 0.0029100137762725353, + 0.004043865948915482, + 0.004177900031208992, + 0.0016064378432929516, + 0.030192099511623383, + 0.023597093299031258, + 0.01930808462202549, + 0.011288877576589584, + 0.027985893189907074, + 0.021747959777712822, + 0.029970571398735046, + 0.007655291352421045, + 0.005360689014196396, + 0.005228283815085888, + 0.0046734558418393135, + 0.004033237230032682, + 0.017679596319794655, + 0.007663182448595762, + 0.014155153185129166, + 0.005123632960021496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.16039389371871948 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906356245279312 + ], + "opposite_logprob_tokens": [ + -2.4640636444091797 + ], + "logprob_reward": -0.08906356245279312, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906356245279312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2571788430213928, + 0.14219316840171814, + 0.08252661675214767, + 0.05915209650993347, + 0.07461214810609818, + 0.08858182281255722, + 0.0504712238907814, + 0.030011195689439774, + 0.1720399260520935, + 0.04323290288448334 + ], + "credit_map": [ + 0.03866513445973396, + 0.009657316841185093, + 0.03704594820737839, + 0.013400700874626637, + 0.035289518535137177, + 0.03536451980471611, + 0.05737018212676048, + 0.030385538935661316, + 0.018782207742333412, + 0.005028198938816786, + 0.025423871353268623, + 0.007899865508079529, + 0.0270378440618515, + 0.014835450798273087, + 0.025255026295781136, + 0.017930712550878525, + 0.01794963888823986, + 0.004307993687689304, + 0.01895013637840748, + 0.006340489722788334, + 0.010159819386899471, + 0.008911105804145336, + 0.009170030243694782, + 0.006737407762557268, + 0.009448779746890068, + 0.0020318273454904556, + 0.011618471704423428, + 0.004313419107347727, + 0.009159857407212257, + 0.006334410049021244, + 0.009125475771725178, + 0.0071198600344359875, + 0.021075772121548653, + 0.0032691939268261194, + 0.020157042890787125, + 0.00533583527430892, + 0.00611203582957387, + 0.006179416552186012, + 0.0066530657932162285, + 0.005829792004078627, + 0.007993632927536964, + 0.0044309161603450775, + 0.015025532804429531, + 0.005593080073595047, + 0.01145418081432581, + 0.017326226457953453, + 0.017128104344010353, + 0.00963014829903841, + 0.00596746988594532, + 0.0030851049814373255, + 0.00940580852329731, + 0.004561733454465866, + 0.0076262387447059155, + 0.007619938347488642, + 0.008013256825506687, + 0.004191672895103693, + 0.004832231439650059, + 0.00235911482013762, + 0.006112587638199329, + 0.002923819934949279, + 0.0038416385650634766, + 0.003854172769933939, + 0.004020668100565672, + 0.002066966611891985, + 0.042147763073444366, + 0.030767489224672318, + 0.02602028101682663, + 0.012154269963502884, + 0.016145821660757065, + 0.02281714417040348, + 0.015724236145615578, + 0.006262937560677528, + 0.0031957863830029964, + 0.003741396823897958, + 0.0056701237335801125, + 0.0040045534260571, + 0.008702684193849564, + 0.0058167558163404465, + 0.00740815931931138, + 0.004693449940532446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.08906356245279312 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-contrast black-and-white close-up, shot from a level angle, focuses on the face of an elderly man, marked by the marks of time. Sharp afternoon sunlight streams through a blinded window, casting a crisscross pattern of light and shadow across his furrowed forehead and sunken cheeks. Bright streaks intertwine with deep shadows, sculpting his features into a richly textured relief. His sparse, short, silver-white hair gleams glaringly in the strong light, while his dark cotton shirt, frayed at the collar, blends almost completely into the darkness of the background. His chest rises and falls slowly and heavily, and then his cracked lips part slightly, releasing a nearly inaudible sigh, profound and filled with endless fatigue. Then, his eyelids, as if carrying a thousand-pound weight, began to droop irresistibly and extremely slowly. The crow's feet at the corners of his eyes were deeply squeezed together with this movement, and finally completely covered his cloudy eyes, blocking out all the light and stories. The whole picture was filled with a profound sense of powerlessness and a quiet atmosphere of sadness in the extreme contrast of light and shadow.", + "seed": 18987, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_5/index_2504_seed_18987_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08977373316884041, + "vlm_logprob": -0.08977373316884041, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.08977373316884041, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052948225289583206 + ], + "opposite_logprob_tokens": [ + -3.552948236465454 + ], + "logprob_reward": -0.052948225289583206, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.052948225289583206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2716730535030365, + 0.09576736390590668, + 0.08966825902462006, + 0.05013875290751457, + 0.07075714319944382, + 0.058476921170949936, + 0.10200513154268265, + 0.05598825961351395, + 0.05666661635041237, + 0.14885851740837097 + ], + "credit_map": [ + 0.04244660958647728, + 0.01993267983198166, + 0.042142126709222794, + 0.03956283628940582, + 0.02379348687827587, + 0.031011341139674187, + 0.035815127193927765, + 0.03696884214878082, + 0.016786061227321625, + 0.014471645466983318, + 0.014262661337852478, + 0.008816772140562534, + 0.00955233909189701, + 0.012369685806334019, + 0.007587561383843422, + 0.01192063931375742, + 0.010759352706372738, + 0.006370809860527515, + 0.013189992867410183, + 0.007106100209057331, + 0.007577172014862299, + 0.012448493391275406, + 0.0113142691552639, + 0.020902063697576523, + 0.010579761117696762, + 0.005941181909292936, + 0.010228783823549747, + 0.003490076633170247, + 0.0038353416603058577, + 0.004560455214232206, + 0.0032165248412638903, + 0.008286627009510994, + 0.009909096173942089, + 0.0033551116939634085, + 0.010124138556420803, + 0.011518893763422966, + 0.00634295167401433, + 0.012296431697905064, + 0.006486350204795599, + 0.01072416640818119, + 0.008126011118292809, + 0.006574966944754124, + 0.0071586440317332745, + 0.005731568206101656, + 0.008195408619940281, + 0.007576076779514551, + 0.005789305083453655, + 0.009324938990175724, + 0.018676845356822014, + 0.013256155885756016, + 0.016274206340312958, + 0.006687924265861511, + 0.007770914118736982, + 0.015887504443526268, + 0.009495762176811695, + 0.013955813832581043, + 0.006807153113186359, + 0.004784178454428911, + 0.007894160225987434, + 0.003174690529704094, + 0.004872417543083429, + 0.008432558737695217, + 0.0055353655479848385, + 0.014487734995782375, + 0.009839728474617004, + 0.008546155877411366, + 0.010114528238773346, + 0.003222796367481351, + 0.004443681798875332, + 0.007084851153194904, + 0.005027235951274633, + 0.008387641049921513, + 0.02438557706773281, + 0.022986752912402153, + 0.022849135100841522, + 0.010203078389167786, + 0.013114890083670616, + 0.018857337534427643, + 0.011973527260124683, + 0.02448822744190693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.052948225289583206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004427164793015 + ], + "opposite_logprob_tokens": [ + -2.6950442790985107 + ], + "logprob_reward": -0.07004427164793015, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004427164793015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24325641989707947, + 0.1349835842847824, + 0.1795898824930191, + 0.1598464846611023, + 0.08053633570671082, + 0.028767718002200127, + 0.059660654515028, + 0.04880431294441223, + 0.030461126938462257, + 0.0340934619307518 + ], + "credit_map": [ + 0.03134949505329132, + 0.01788414642214775, + 0.03607536107301712, + 0.02224021404981613, + 0.030125869438052177, + 0.039620645344257355, + 0.02712412178516388, + 0.03883657976984978, + 0.015361190773546696, + 0.010151553899049759, + 0.018117237836122513, + 0.012150810100138187, + 0.01874115690588951, + 0.022167887538671494, + 0.014123212546110153, + 0.02417055144906044, + 0.020587284117937088, + 0.009065198712050915, + 0.026162302121520042, + 0.014139289036393166, + 0.016019104048609734, + 0.025468973442912102, + 0.023487770929932594, + 0.04465998709201813, + 0.019870484247803688, + 0.008243921212852001, + 0.02378806285560131, + 0.016788765788078308, + 0.012167856097221375, + 0.02194288931787014, + 0.011880822479724884, + 0.045163705945014954, + 0.011997676454484463, + 0.004387723747640848, + 0.011767733842134476, + 0.012876329012215137, + 0.008214130066335201, + 0.014005971141159534, + 0.005243695806711912, + 0.012043081223964691, + 0.00402590399608016, + 0.002462688134983182, + 0.0036602721083909273, + 0.0012918856227770448, + 0.003191846888512373, + 0.00449081975966692, + 0.0036253242287784815, + 0.006018978077918291, + 0.009233567863702774, + 0.007177630439400673, + 0.009253828786313534, + 0.0033870136830955744, + 0.005852595902979374, + 0.009324096143245697, + 0.004881961736828089, + 0.010549960657954216, + 0.006086005363613367, + 0.002948591485619545, + 0.007722990121692419, + 0.0027245907112956047, + 0.003994794096797705, + 0.00630924804136157, + 0.004355651792138815, + 0.014662445522844791, + 0.0043869721703231335, + 0.001870261738076806, + 0.005257421638816595, + 0.001499425619840622, + 0.0031210260931402445, + 0.005169413518160582, + 0.002707810839638114, + 0.006448797881603241, + 0.003914773929864168, + 0.0026022931560873985, + 0.004247245378792286, + 0.003045099088922143, + 0.003988408483564854, + 0.005530992057174444, + 0.003985770046710968, + 0.006778881885111332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.07004427164793015 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07007816433906555 + ], + "opposite_logprob_tokens": [ + -2.695078134536743 + ], + "logprob_reward": -0.07007816433906555, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07007816433906555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20456334948539734, + 0.14076180756092072, + 0.09972099214792252, + 0.13904814422130585, + 0.07572170346975327, + 0.04203382879495621, + 0.15317420661449432, + 0.04293012246489525, + 0.03197295591235161, + 0.07007285207509995 + ], + "credit_map": [ + 0.03476666286587715, + 0.02320038340985775, + 0.036036841571331024, + 0.018523819744586945, + 0.016713595017790794, + 0.02743373066186905, + 0.0173227246850729, + 0.03056558594107628, + 0.022192563861608505, + 0.01698993891477585, + 0.019199656322598457, + 0.017865872010588646, + 0.011916582472622395, + 0.016345953568816185, + 0.015450910665094852, + 0.02080032229423523, + 0.013043106533586979, + 0.008368569426238537, + 0.016421806067228317, + 0.009839820675551891, + 0.00546357361599803, + 0.013812349177896976, + 0.010010994970798492, + 0.022760774940252304, + 0.021034669131040573, + 0.01276851911097765, + 0.02219671569764614, + 0.009989376179873943, + 0.010323869995772839, + 0.01721198670566082, + 0.00936499796807766, + 0.036158014088869095, + 0.010085182264447212, + 0.004416594281792641, + 0.011235239915549755, + 0.01090120617300272, + 0.006069527938961983, + 0.011493770405650139, + 0.007142414804548025, + 0.01437776442617178, + 0.005233129486441612, + 0.00517412880435586, + 0.00608346750959754, + 0.003949248231947422, + 0.004761372692883015, + 0.00679089967161417, + 0.003958623856306076, + 0.0060829599387943745, + 0.027568338438868523, + 0.0278775691986084, + 0.023460356518626213, + 0.006878627929836512, + 0.011810083873569965, + 0.024352267384529114, + 0.011778008192777634, + 0.01944895274937153, + 0.0068660518154501915, + 0.004488502629101276, + 0.007447709329426289, + 0.002632610499858856, + 0.003185159293934703, + 0.005932945758104324, + 0.0038157703820616007, + 0.008561373688280582, + 0.005535558797419071, + 0.004278555978089571, + 0.005221860483288765, + 0.0027998904697597027, + 0.002902531996369362, + 0.004031320568174124, + 0.00230404338799417, + 0.004899194464087486, + 0.009056778624653816, + 0.009359653107821941, + 0.010157184675335884, + 0.004806556273251772, + 0.008167834021151066, + 0.010509018786251545, + 0.008184841834008694, + 0.009830979630351067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.07007816433906555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428518146276474 + ], + "opposite_logprob_tokens": [ + -2.0178518295288086 + ], + "logprob_reward": -0.1428518146276474, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428518146276474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16193440556526184, + 0.10227327793836594, + 0.15230336785316467, + 0.16556185483932495, + 0.08766920864582062, + 0.09056861698627472, + 0.11629407852888107, + 0.05644569173455238, + 0.036028992384672165, + 0.030920526012778282 + ], + "credit_map": [ + 0.03250706568360329, + 0.01331248227506876, + 0.028251489624381065, + 0.02253701351583004, + 0.01432887464761734, + 0.019695362076163292, + 0.013009943999350071, + 0.018292153254151344, + 0.014630777761340141, + 0.009133287705481052, + 0.013859552331268787, + 0.012598395347595215, + 0.009529135189950466, + 0.013625901192426682, + 0.008662363514304161, + 0.02023385651409626, + 0.02389495074748993, + 0.007792335934937, + 0.023450367152690887, + 0.014321578666567802, + 0.011350688524544239, + 0.020477432757616043, + 0.017859147861599922, + 0.0331568568944931, + 0.024334127083420753, + 0.011237706057727337, + 0.0230999868363142, + 0.021636707708239555, + 0.013628033921122551, + 0.020651428028941154, + 0.012602750211954117, + 0.038371093571186066, + 0.013608093373477459, + 0.008019930683076382, + 0.01370642427355051, + 0.007859461940824986, + 0.007691296748816967, + 0.012736954726278782, + 0.00637472327798605, + 0.01767231523990631, + 0.01481598149985075, + 0.008047882467508316, + 0.014532089233398438, + 0.005810450296849012, + 0.009413504973053932, + 0.009883979335427284, + 0.007592386100441217, + 0.020472338423132896, + 0.015609237365424633, + 0.012722210958600044, + 0.01658172719180584, + 0.0059548914432525635, + 0.012427283450961113, + 0.02501816116273403, + 0.009110364131629467, + 0.018870191648602486, + 0.007460208144038916, + 0.004104268737137318, + 0.008425487205386162, + 0.0057669878005981445, + 0.0060982308350503445, + 0.008760436438024044, + 0.007837818935513496, + 0.007992252707481384, + 0.004250338766723871, + 0.0035975908394902945, + 0.005166835617274046, + 0.0032016148325055838, + 0.00302034430205822, + 0.006848025601357222, + 0.0035401552449911833, + 0.006404082756489515, + 0.004035984631627798, + 0.003172080498188734, + 0.004227536730468273, + 0.0022883072961121798, + 0.0023587069008499384, + 0.006188726983964443, + 0.003287149593234062, + 0.005362029653042555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.1428518146276474 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294618993997574 + ], + "opposite_logprob_tokens": [ + -2.2379462718963623 + ], + "logprob_reward": -0.11294618993997574, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294618993997574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13661131262779236, + 0.09029325842857361, + 0.0502166673541069, + 0.280832439661026, + 0.15214940905570984, + 0.08242934942245483, + 0.05155609920620918, + 0.031170355156064034, + 0.04696260765194893, + 0.07777849584817886 + ], + "credit_map": [ + 0.02141484059393406, + 0.008720031008124352, + 0.022722292691469193, + 0.015187719836831093, + 0.012899436056613922, + 0.01717144250869751, + 0.01558616105467081, + 0.022909393534064293, + 0.011667474173009396, + 0.007109424564987421, + 0.011473345570266247, + 0.013382306322455406, + 0.010292675346136093, + 0.014044632203876972, + 0.008351190946996212, + 0.013972211629152298, + 0.007121250964701176, + 0.0029366284143179655, + 0.007918309420347214, + 0.0030009697657078505, + 0.005918706301599741, + 0.007886531762778759, + 0.005605045240372419, + 0.009829225018620491, + 0.0382220521569252, + 0.015357057563960552, + 0.04489586874842644, + 0.02262824960052967, + 0.021961623802781105, + 0.03903719782829285, + 0.02101251669228077, + 0.07771787792444229, + 0.023670032620429993, + 0.007989632897078991, + 0.02398226223886013, + 0.01906253769993782, + 0.014462706632912159, + 0.028938205912709236, + 0.011167870834469795, + 0.02287616766989231, + 0.012604737654328346, + 0.0051771411672234535, + 0.01124898623675108, + 0.009222025983035564, + 0.009581055492162704, + 0.013154879212379456, + 0.008121795952320099, + 0.01331872958689928, + 0.00768863083794713, + 0.006718664430081844, + 0.007491145748645067, + 0.0029372654389590025, + 0.005416783969849348, + 0.009102527983486652, + 0.004166032187640667, + 0.008035046979784966, + 0.0033806688152253628, + 0.0020145687740296125, + 0.004955158568918705, + 0.003746881615370512, + 0.0030413437634706497, + 0.0043903919868171215, + 0.004395077936351299, + 0.005246262066066265, + 0.005517751444131136, + 0.004592574667185545, + 0.006118881516158581, + 0.005296892952173948, + 0.0032872981391847134, + 0.007908660918474197, + 0.0046051545068621635, + 0.009635395370423794, + 0.008364303037524223, + 0.004189826548099518, + 0.010135067626833916, + 0.00884857028722763, + 0.008573842234909534, + 0.012493491172790527, + 0.008533927612006664, + 0.01663947105407715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.11294618993997574 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-contrast black-and-white close-up, shot from a level angle, focuses on the face of an elderly man, marked by the marks of time. Sharp afternoon sunlight streams through a blinded window, casting a crisscross pattern of light and shadow across his furrowed forehead and sunken cheeks. Bright streaks intertwine with deep shadows, sculpting his features into a richly textured relief. His sparse, short, silver-white hair gleams glaringly in the strong light, while his dark cotton shirt, frayed at the collar, blends almost completely into the darkness of the background. His chest rises and falls slowly and heavily, and then his cracked lips part slightly, releasing a nearly inaudible sigh, profound and filled with endless fatigue. Then, his eyelids, as if carrying a thousand-pound weight, began to droop irresistibly and extremely slowly. The crow's feet at the corners of his eyes were deeply squeezed together with this movement, and finally completely covered his cloudy eyes, blocking out all the light and stories. The whole picture was filled with a profound sense of powerlessness and a quiet atmosphere of sadness in the extreme contrast of light and shadow.", + "seed": 18988, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_5/index_2504_seed_18988_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06574963703751564, + "vlm_logprob": -0.06574963703751564, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.06574963703751564, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.057216446846723557 + ], + "opposite_logprob_tokens": [ + -4.307216644287109 + ], + "logprob_reward": -0.057216446846723557, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.057216446846723557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2879842221736908, + 0.08094896376132965, + 0.06824775040149689, + 0.11185069382190704, + 0.07096506655216217, + 0.12259836494922638, + 0.053041066974401474, + 0.04953523725271225, + 0.05683236941695213, + 0.097996287047863 + ], + "credit_map": [ + 0.022734057158231735, + 0.0303233340382576, + 0.042762383818626404, + 0.048243213444948196, + 0.04192794859409332, + 0.037709735333919525, + 0.030470719560980797, + 0.03381281718611717, + 0.0090000806376338, + 0.010920198634266853, + 0.0077527971006929874, + 0.01171449851244688, + 0.013514443300664425, + 0.007725819479674101, + 0.013658052310347557, + 0.006663077976554632, + 0.00733537832275033, + 0.006861909758299589, + 0.008853009901940823, + 0.005113403778523207, + 0.0071630035527050495, + 0.01025976613163948, + 0.011586139909923077, + 0.011075139045715332, + 0.017943304032087326, + 0.015762776136398315, + 0.010533744469285011, + 0.01042869035154581, + 0.01260730717331171, + 0.013013744726777077, + 0.01817328855395317, + 0.013387837447226048, + 0.01451126579195261, + 0.006327812559902668, + 0.01006067730486393, + 0.007145206909626722, + 0.005511481314897537, + 0.008537320420145988, + 0.010364968329668045, + 0.008506334386765957, + 0.016222424805164337, + 0.01721574366092682, + 0.009156758897006512, + 0.015298016369342804, + 0.014439700171351433, + 0.01642247848212719, + 0.014557878486812115, + 0.019285358488559723, + 0.008832134306430817, + 0.006123017985373735, + 0.005443754605948925, + 0.004843550268560648, + 0.008350723423063755, + 0.0071221101097762585, + 0.005454848986119032, + 0.006870931014418602, + 0.0056089130230247974, + 0.006272325292229652, + 0.007439290639013052, + 0.0053626918233931065, + 0.0048926640301942825, + 0.007953539490699768, + 0.005108133889734745, + 0.006897678133100271, + 0.009421215392649174, + 0.005981819704174995, + 0.008758815005421638, + 0.004465681035071611, + 0.005826877895742655, + 0.007971676997840405, + 0.00708276079967618, + 0.0073235216550529, + 0.013136906549334526, + 0.00928947702050209, + 0.010951165109872818, + 0.010517655871808529, + 0.01134981494396925, + 0.015681646764278412, + 0.00892651453614235, + 0.018143100664019585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.057216446846723557 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637951821088791 + ], + "opposite_logprob_tokens": [ + -3.6513795852661133 + ], + "logprob_reward": -0.02637951821088791, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637951821088791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25697118043899536, + 0.08064796030521393, + 0.06106987223029137, + 0.07813116163015366, + 0.05605517327785492, + 0.314336359500885, + 0.028219174593687057, + 0.03498094528913498, + 0.03138790652155876, + 0.05820024013519287 + ], + "credit_map": [ + 0.03609773889183998, + 0.03744453191757202, + 0.025736438110470772, + 0.035032935440540314, + 0.02873675897717476, + 0.02813797816634178, + 0.03445591777563095, + 0.03132887929677963, + 0.012515531852841377, + 0.010953797027468681, + 0.00852200761437416, + 0.007260849699378014, + 0.010993586853146553, + 0.010619031265377998, + 0.008653280325233936, + 0.011129873804748058, + 0.009091533720493317, + 0.00829373300075531, + 0.0063499631360173225, + 0.004443574231117964, + 0.006405968219041824, + 0.009027866646647453, + 0.007254352793097496, + 0.01020287536084652, + 0.009906698949635029, + 0.011075970716774464, + 0.00780725572258234, + 0.0077935256995260715, + 0.011305070482194424, + 0.010610722005367279, + 0.010231303051114082, + 0.009400610812008381, + 0.006371369585394859, + 0.0069297500886023045, + 0.008111129514873028, + 0.006630840245634317, + 0.005512718576937914, + 0.005674951244145632, + 0.010732976719737053, + 0.0060914382338523865, + 0.027546564117074013, + 0.044374801218509674, + 0.030842315405607224, + 0.07901682704687119, + 0.027364730834960938, + 0.0335547961294651, + 0.035443492233753204, + 0.036192841827869415, + 0.0037691036704927683, + 0.004057900980114937, + 0.00384186627343297, + 0.002368355169892311, + 0.00431792251765728, + 0.0036219903267920017, + 0.0033154997508972883, + 0.0029265359044075012, + 0.004661732353270054, + 0.004086094442754984, + 0.004923871252685785, + 0.004302071873098612, + 0.003060552990064025, + 0.004940086510032415, + 0.003442878369241953, + 0.005563658196479082, + 0.0033127861097455025, + 0.002925766631960869, + 0.0038132553454488516, + 0.0024725564289838076, + 0.003924539778381586, + 0.005015952978283167, + 0.003368745557963848, + 0.006554306019097567, + 0.007771978620439768, + 0.007289687637239695, + 0.011373779736459255, + 0.007582563441246748, + 0.004282976035028696, + 0.007793244440108538, + 0.00594803923740983, + 0.006157974246889353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.02637951821088791 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311820864677429 + ], + "opposite_logprob_tokens": [ + -3.1681182384490967 + ], + "logprob_reward": -0.04311820864677429, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311820864677429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16761897504329681, + 0.09688060730695724, + 0.08318765461444855, + 0.16879534721374512, + 0.20712734758853912, + 0.1032329574227333, + 0.04205530881881714, + 0.023390432819724083, + 0.05152781680226326, + 0.05618356540799141 + ], + "credit_map": [ + 0.02819421887397766, + 0.024045204743742943, + 0.015267938375473022, + 0.010267456993460655, + 0.01969943568110466, + 0.02179105021059513, + 0.024180499836802483, + 0.024173179641366005, + 0.013052736409008503, + 0.01384686678647995, + 0.013909121043980122, + 0.008627998642623425, + 0.009343376383185387, + 0.013312360271811485, + 0.013958186842501163, + 0.010829959064722061, + 0.01548716239631176, + 0.008049647323787212, + 0.012436944991350174, + 0.00503890635445714, + 0.009977496229112148, + 0.011876458302140236, + 0.010030402801930904, + 0.010290632024407387, + 0.04035574942827225, + 0.0216361191123724, + 0.015209980309009552, + 0.009310487657785416, + 0.017225589603185654, + 0.021323135122656822, + 0.022588875144720078, + 0.021145416423678398, + 0.03569286689162254, + 0.02574465610086918, + 0.021244391798973083, + 0.014499946497380733, + 0.0196703989058733, + 0.026165250688791275, + 0.035290513187646866, + 0.02881932258605957, + 0.010507827624678612, + 0.01496539730578661, + 0.007746404968202114, + 0.01699763722717762, + 0.014665995724499226, + 0.010402332991361618, + 0.01902717910706997, + 0.008920185267925262, + 0.005862660706043243, + 0.007994276471436024, + 0.0033469381742179394, + 0.0044388663955032825, + 0.0074561103247106075, + 0.004481788258999586, + 0.003996802028268576, + 0.00447786645963788, + 0.002157295122742653, + 0.0021628784015774727, + 0.004112746566534042, + 0.0019483587238937616, + 0.0029923049733042717, + 0.003549693152308464, + 0.0037627345882356167, + 0.00270442315377295, + 0.0053975763730704784, + 0.00522691709920764, + 0.00807843916118145, + 0.0034776334650814533, + 0.007511971518397331, + 0.007431903854012489, + 0.004288372118026018, + 0.0101150032132864, + 0.004691922105848789, + 0.004722713492810726, + 0.00801241397857666, + 0.00807492807507515, + 0.005188785959035158, + 0.008926928043365479, + 0.006559000816196203, + 0.010006872937083244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.04311820864677429 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296652257442474 + ], + "opposite_logprob_tokens": [ + -2.237966537475586 + ], + "logprob_reward": -0.11296652257442474, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296652257442474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1370045393705368, + 0.0758666917681694, + 0.07407310605049133, + 0.14254041016101837, + 0.07391440123319626, + 0.0856940746307373, + 0.05821915343403816, + 0.10150536149740219, + 0.1771593540906906, + 0.07402286678552628 + ], + "credit_map": [ + 0.014911591075360775, + 0.015673242509365082, + 0.014157667756080627, + 0.012014398351311684, + 0.018983326852321625, + 0.01841355301439762, + 0.023323936387896538, + 0.019526824355125427, + 0.006391366012394428, + 0.008531838655471802, + 0.008049785159528255, + 0.007301815785467625, + 0.009138335473835468, + 0.010726610198616982, + 0.01413523219525814, + 0.011591707356274128, + 0.01060289703309536, + 0.010440372861921787, + 0.008000867441296577, + 0.008027935400605202, + 0.00779257295653224, + 0.009904359467327595, + 0.008808587677776814, + 0.010495506227016449, + 0.020074304193258286, + 0.022908160462975502, + 0.014299731701612473, + 0.01405961811542511, + 0.01612026058137417, + 0.018998173996806145, + 0.01653858833014965, + 0.019541580229997635, + 0.011154976673424244, + 0.006776425521820784, + 0.008894147351384163, + 0.005761910695582628, + 0.0072577642276883125, + 0.010886073112487793, + 0.010045165196061134, + 0.013137932866811752, + 0.010556328110396862, + 0.012056030333042145, + 0.01075895968824625, + 0.006997620686888695, + 0.011883088387548923, + 0.010290026664733887, + 0.013716194778680801, + 0.00943582784384489, + 0.007480195257812738, + 0.00871103536337614, + 0.007694105617702007, + 0.005415256135165691, + 0.010190497152507305, + 0.006892194040119648, + 0.00613612774759531, + 0.005699743051081896, + 0.01107566524296999, + 0.010724608786404133, + 0.016925884410738945, + 0.0075439452193677425, + 0.012791290879249573, + 0.01570615917444229, + 0.01704467087984085, + 0.009693140164017677, + 0.016270531341433525, + 0.014226069673895836, + 0.02771136537194252, + 0.013951416127383709, + 0.026026058942079544, + 0.026284899562597275, + 0.01891598291695118, + 0.0337730348110199, + 0.0062698801048099995, + 0.007403428200632334, + 0.008693424053490162, + 0.006848487537354231, + 0.010616958141326904, + 0.011479408480226994, + 0.011368430219590664, + 0.011342845857143402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.11296652257442474 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video a high-contrast black-and-white close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject an elderly man with sparse, short, silver-white hair and a dark cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does sharp afternoon sunlight stream through a window, casting a crisscross pattern of light and shadow on his face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the man's eyelids slowly droop to cover his eyes, accompanied by the appearance of deep crow's feet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a level angle with the man's chest rising and falling slowly and heavily?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890674889087677 + ], + "opposite_logprob_tokens": [ + -2.4640674591064453 + ], + "logprob_reward": -0.0890674889087677, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890674889087677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21368028223514557, + 0.10786335170269012, + 0.16157598793506622, + 0.05936099961400032, + 0.1036277636885643, + 0.12026132643222809, + 0.06673707813024521, + 0.055673837661743164, + 0.05307246744632721, + 0.05814690887928009 + ], + "credit_map": [ + 0.024213459342718124, + 0.029674554243683815, + 0.02203063666820526, + 0.03926929831504822, + 0.021761223673820496, + 0.022866463288664818, + 0.032585259526968, + 0.021279381588101387, + 0.010172377340495586, + 0.0163270253688097, + 0.01227623876184225, + 0.011674491688609123, + 0.015124836936593056, + 0.013608898036181927, + 0.016736825928092003, + 0.011942654848098755, + 0.027969317510724068, + 0.018917953595519066, + 0.016320740804076195, + 0.012378928251564503, + 0.017725560814142227, + 0.022931955754756927, + 0.020349061116576195, + 0.02498246729373932, + 0.013595033437013626, + 0.0060531971976161, + 0.004651086404919624, + 0.00333013036288321, + 0.00662821251899004, + 0.008608781732618809, + 0.007147275377064943, + 0.009347286075353622, + 0.01146833598613739, + 0.014501559548079967, + 0.00942257884889841, + 0.008527376689016819, + 0.012879127636551857, + 0.011929512023925781, + 0.021561583504080772, + 0.013337685726583004, + 0.013392006978392601, + 0.015237554907798767, + 0.014468364417552948, + 0.009540471248328686, + 0.02093062363564968, + 0.014638184569776058, + 0.01744733192026615, + 0.01460678968578577, + 0.010395546443760395, + 0.009393775835633278, + 0.009031455963850021, + 0.00635701697319746, + 0.00915488600730896, + 0.008025307208299637, + 0.005503963213413954, + 0.008875129744410515, + 0.006943318992853165, + 0.007927050814032555, + 0.0064759571105241776, + 0.0054182675667107105, + 0.0071630594320595264, + 0.007576898206025362, + 0.008038514293730259, + 0.006130772177129984, + 0.007425932213664055, + 0.006920287851244211, + 0.006345164962112904, + 0.00607506325468421, + 0.005648083984851837, + 0.006574531551450491, + 0.007237580139189959, + 0.006845818366855383, + 0.004368667025119066, + 0.006538085639476776, + 0.006322277709841728, + 0.0061224643141031265, + 0.009927788749337196, + 0.008754407055675983, + 0.007957104593515396, + 0.008156114257872105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0890674889087677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_6/index_tensor([1651, 1651])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_6/index_tensor([1651, 1651])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..26e6e309385564889043d722748417bd22d93387 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_6/index_tensor([1651, 1651])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, low-angle shot, rendered with photorealistic CGI cinematic realism, slowly advances onto the surface of a violent, roiling alien sea, dominated by a massive, angular Hunter-Warrior starship. Mountains of viscous, inky-black water undulate, their crests exploding into chaotic foam beneath a sky choked by scarred, dark gray storm clouds intermittently illuminated by silent, branching bursts of purple lightning. The starship's black hull, a masterpiece of hostile geometry, constructed from a matte, non-reflective composite, relentlessly slices through towering waves, sending volcanic spray blasting into the air. Dramatic lighting cast by strobe lightning captures the atomized spray and foam, causing them to shimmer like shattered glass against the vessel's obsidian form and create a stark, shifting black-and-white contrast effect, momentarily revealing the intricate texture of its armor plating. The ship's immense scale is palpable, its unstoppable forward momentum contrasting sharply with the raw, chaotic force of the ocean, creating an atmosphere of tension and awe-inspiring, overwhelming power.", + "seed": 11284, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_6/index_1651_seed_11284_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11209977427497506, + "vlm_logprob": -0.11209977427497506, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.11209977427497506, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0899805873632431 + ], + "opposite_logprob_tokens": [ + -2.7149806022644043 + ], + "logprob_reward": -0.0899805873632431, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0899805873632431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21461829543113708, + 0.042742472141981125, + 0.050099004060029984, + 0.08654838055372238, + 0.049575336277484894, + 0.052861109375953674, + 0.080010786652565, + 0.15976136922836304, + 0.1735527217388153, + 0.0902305468916893 + ], + "credit_map": [ + 0.023232553154230118, + 0.02712363190948963, + 0.04453463852405548, + 0.024109944701194763, + 0.030579188838601112, + 0.028907716274261475, + 0.015706341713666916, + 0.02042427286505699, + 0.00362578546628356, + 0.004516839515417814, + 0.006815759930759668, + 0.004986501298844814, + 0.008516451343894005, + 0.006553007289767265, + 0.003246922045946121, + 0.004481203854084015, + 0.004448019899427891, + 0.0056680431589484215, + 0.007436729967594147, + 0.006911813281476498, + 0.00677674962207675, + 0.007673813495784998, + 0.005430529359728098, + 0.005753303878009319, + 0.009195618331432343, + 0.011005127802491188, + 0.008715636096894741, + 0.010364987887442112, + 0.01637093350291252, + 0.013129289261996746, + 0.008040759712457657, + 0.009726027026772499, + 0.007244178093969822, + 0.005377256311476231, + 0.0036061936989426613, + 0.005925774574279785, + 0.009424126707017422, + 0.007166921626776457, + 0.005354402586817741, + 0.005476484540849924, + 0.003512308467179537, + 0.006942268926650286, + 0.0027998846489936113, + 0.005735528655350208, + 0.009999026544392109, + 0.013109119608998299, + 0.0067057437263429165, + 0.004057231359183788, + 0.005722064524888992, + 0.01179538108408451, + 0.00809888169169426, + 0.008024366572499275, + 0.013096695765852928, + 0.015640398487448692, + 0.00960599910467863, + 0.008026997558772564, + 0.011113735847175121, + 0.01984427310526371, + 0.029134700074791908, + 0.018531668931245804, + 0.022234030067920685, + 0.025431202724575996, + 0.01907208189368248, + 0.014399677515029907, + 0.01185606699436903, + 0.02295694500207901, + 0.026243917644023895, + 0.03607972338795662, + 0.017926407977938652, + 0.019752709195017815, + 0.021030249074101448, + 0.01770670711994171, + 0.012127377092838287, + 0.01706553064286709, + 0.010535378940403461, + 0.009187547490000725, + 0.013677588663995266, + 0.012789447791874409, + 0.006692797876894474, + 0.008154882118105888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0899805873632431 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062089066952466965 + ], + "opposite_logprob_tokens": [ + -2.812088966369629 + ], + "logprob_reward": -0.062089066952466965, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062089066952466965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13923580944538116, + 0.06461665779352188, + 0.04329092428088188, + 0.08107681572437286, + 0.05188501998782158, + 0.055714044719934464, + 0.08136327564716339, + 0.09177501499652863, + 0.2106974571943283, + 0.18034499883651733 + ], + "credit_map": [ + 0.014127785339951515, + 0.013708997517824173, + 0.023264190182089806, + 0.017514871433377266, + 0.019598882645368576, + 0.02138264663517475, + 0.014669448137283325, + 0.01496899500489235, + 0.0061395433731377125, + 0.005400403402745724, + 0.010601832531392574, + 0.008699397556483746, + 0.009730241261422634, + 0.008958907797932625, + 0.007550208829343319, + 0.007536121644079685, + 0.003938153386116028, + 0.006302946712821722, + 0.005500072613358498, + 0.004444614984095097, + 0.005754904821515083, + 0.006020511966198683, + 0.005334843881428242, + 0.005994874518364668, + 0.009860754013061523, + 0.01328018493950367, + 0.009319799020886421, + 0.00852985680103302, + 0.01113604661077261, + 0.011264371685683727, + 0.008526368997991085, + 0.009159432724118233, + 0.004724707454442978, + 0.005728565622121096, + 0.0037064452189952135, + 0.006226697470992804, + 0.010415418073534966, + 0.008049129508435726, + 0.006418740376830101, + 0.006615315563976765, + 0.005525590851902962, + 0.008603150956332684, + 0.0038835867308080196, + 0.006839462090283632, + 0.006986423395574093, + 0.008151066489517689, + 0.007393455598503351, + 0.008331310003995895, + 0.005979776382446289, + 0.011510713025927544, + 0.008218809962272644, + 0.007795772515237331, + 0.016596602275967598, + 0.012316159904003143, + 0.008765948936343193, + 0.010179501958191395, + 0.008719924837350845, + 0.011576210148632526, + 0.012054326944053173, + 0.008264700882136822, + 0.014741238206624985, + 0.015000749379396439, + 0.009925241582095623, + 0.011492620222270489, + 0.011823322623968124, + 0.02323313243687153, + 0.03135082870721817, + 0.06197388097643852, + 0.018188754096627235, + 0.02028563991189003, + 0.022721106186509132, + 0.021120792254805565, + 0.01826985366642475, + 0.028860507532954216, + 0.02164916880428791, + 0.01697540283203125, + 0.033006392419338226, + 0.025815486907958984, + 0.01932862587273121, + 0.016439564526081085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.062089066952466965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606742851436138 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606742851436138, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606742851436138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062131673097610474, + 0.036492835730314255, + 0.02766786888241768, + 0.038223911076784134, + 0.03323002904653549, + 0.02328375354409218, + 0.05101291835308075, + 0.25848695635795593, + 0.1963701844215393, + 0.2730998992919922 + ], + "credit_map": [ + 0.006051626056432724, + 0.006642198655754328, + 0.014458457008004189, + 0.006631971802562475, + 0.00846595223993063, + 0.008958729915320873, + 0.0063118659891188145, + 0.004610870033502579, + 0.0053839911706745625, + 0.0035300608724355698, + 0.004835411906242371, + 0.0034577660262584686, + 0.005481142085045576, + 0.004481832962483168, + 0.005402599927037954, + 0.003920032177120447, + 0.004005720838904381, + 0.003379200818017125, + 0.003997945226728916, + 0.003446259768679738, + 0.0033887478057295084, + 0.0035876459442079067, + 0.00310313212685287, + 0.0027592158876359463, + 0.004208684433251619, + 0.0051913028582930565, + 0.003654383821412921, + 0.006292845122516155, + 0.005255253054201603, + 0.005446679424494505, + 0.003881445387378335, + 0.004293317906558514, + 0.005121168680489063, + 0.0025289871264249086, + 0.0026091455947607756, + 0.003744947724044323, + 0.007644264493137598, + 0.0047468156553804874, + 0.003304135287180543, + 0.0035305635537952185, + 0.0014634702820330858, + 0.0028311521746218204, + 0.0015553280245512724, + 0.0025903922505676746, + 0.00394618883728981, + 0.0053940643556416035, + 0.002668113214895129, + 0.002835044637322426, + 0.003921451512724161, + 0.006090667098760605, + 0.00439395010471344, + 0.004835287109017372, + 0.010963286273181438, + 0.009264016523957253, + 0.006148942746222019, + 0.005395315121859312, + 0.02751978486776352, + 0.03195685148239136, + 0.05762089416384697, + 0.03598589822649956, + 0.019930684939026833, + 0.03785545006394386, + 0.024249903857707977, + 0.023367486894130707, + 0.015871180221438408, + 0.01973862759768963, + 0.034104324877262115, + 0.04500285163521767, + 0.019540153443813324, + 0.02126535028219223, + 0.023850319907069206, + 0.016997378319501877, + 0.02685236930847168, + 0.03774704784154892, + 0.04127662628889084, + 0.047864802181720734, + 0.03575412184000015, + 0.03692074865102768, + 0.022252341732382774, + 0.02443184331059456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.01606742851436138 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38710716366767883 + ], + "opposite_logprob_tokens": [ + -1.1371071338653564 + ], + "logprob_reward": -0.38710716366767883, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38710716366767883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1358969658613205, + 0.0886961966753006, + 0.05962495505809784, + 0.0879383310675621, + 0.03610283508896828, + 0.032228171825408936, + 0.03198976814746857, + 0.15233494341373444, + 0.193670392036438, + 0.18151740729808807 + ], + "credit_map": [ + 0.017621032893657684, + 0.017323486506938934, + 0.014901048503816128, + 0.015983648598194122, + 0.02322872169315815, + 0.02189309149980545, + 0.010778353549540043, + 0.014167586341500282, + 0.01020270399749279, + 0.010482629761099815, + 0.01179011631757021, + 0.008708578534424305, + 0.014598935842514038, + 0.01454765535891056, + 0.00918434839695692, + 0.009181232191622257, + 0.006060406100004911, + 0.007526266388595104, + 0.007362234406173229, + 0.005455421283841133, + 0.010466068983078003, + 0.010144044645130634, + 0.004998228512704372, + 0.007612285204231739, + 0.011847708374261856, + 0.009220588952302933, + 0.007132064551115036, + 0.00862209778279066, + 0.012534404173493385, + 0.01333770714700222, + 0.012819359079003334, + 0.012424408458173275, + 0.002963343635201454, + 0.004480205476284027, + 0.0036065340973436832, + 0.004614093340933323, + 0.006694391369819641, + 0.005342690274119377, + 0.0044594574719667435, + 0.003942121751606464, + 0.002821270376443863, + 0.004014994949102402, + 0.001925271120853722, + 0.003774141427129507, + 0.006828807760030031, + 0.005600498057901859, + 0.0036856152582913637, + 0.0035775748547166586, + 0.003099829191341996, + 0.004321117885410786, + 0.003618778195232153, + 0.004008572082966566, + 0.004027452785521746, + 0.005325254984200001, + 0.004154436755925417, + 0.0034343302249908447, + 0.016807325184345245, + 0.023338431492447853, + 0.026252232491970062, + 0.02176373079419136, + 0.012466314248740673, + 0.020414436236023903, + 0.015663964673876762, + 0.015628507360816002, + 0.014182980172336102, + 0.021594122052192688, + 0.030373109504580498, + 0.03553086519241333, + 0.031658515334129333, + 0.024760523810982704, + 0.017811493948101997, + 0.017758809030056, + 0.025079520419239998, + 0.03332376107573509, + 0.021784214302897453, + 0.017728403210639954, + 0.025356002151966095, + 0.02291673608124256, + 0.016752472147345543, + 0.01857631467282772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.38710716366767883 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254624877125025 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254624877125025, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254624877125025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1687493920326233, + 0.04397426173090935, + 0.04247143864631653, + 0.04597114399075508, + 0.053748227655887604, + 0.10727795958518982, + 0.11530929803848267, + 0.17955829203128815, + 0.16800875961780548, + 0.07493116706609726 + ], + "credit_map": [ + 0.026951845735311508, + 0.015634646639227867, + 0.022246982902288437, + 0.0159863643348217, + 0.029169702902436256, + 0.024370944127440453, + 0.01738581620156765, + 0.017003094777464867, + 0.002925216220319271, + 0.005157929379492998, + 0.00553317042067647, + 0.0044726780615746975, + 0.010168151929974556, + 0.006923312321305275, + 0.004838315770030022, + 0.003955489490181208, + 0.004521530587226152, + 0.005951059050858021, + 0.005535452626645565, + 0.003708046628162265, + 0.0077326977625489235, + 0.006924303248524666, + 0.0036270387936383486, + 0.004471314139664173, + 0.0038587874732911587, + 0.006555563770234585, + 0.003248798195272684, + 0.004188357852399349, + 0.008198501542210579, + 0.0077856010757386684, + 0.006564934737980366, + 0.005570597015321255, + 0.006946929730474949, + 0.005735950078815222, + 0.005066810641437769, + 0.00555680925026536, + 0.012626735493540764, + 0.007818865589797497, + 0.0046449038200080395, + 0.00535122025758028, + 0.01171724870800972, + 0.016050899401307106, + 0.007318820338696241, + 0.010076506994664669, + 0.01741175912320614, + 0.01727946288883686, + 0.013820573687553406, + 0.013602692633867264, + 0.009012583643198013, + 0.020181195810437202, + 0.008917523548007011, + 0.007741877343505621, + 0.02966618537902832, + 0.019948801025748253, + 0.010109768249094486, + 0.009731363505125046, + 0.019413519650697708, + 0.017771195620298386, + 0.028485557064414024, + 0.017122391611337662, + 0.03075135499238968, + 0.029691556468605995, + 0.01750047318637371, + 0.018822243437170982, + 0.016502514481544495, + 0.02306767925620079, + 0.02464286983013153, + 0.01170168723911047, + 0.042618393898010254, + 0.02152138203382492, + 0.013456873595714569, + 0.014497369527816772, + 0.006631400901824236, + 0.013682039454579353, + 0.007864866405725479, + 0.009549865499138832, + 0.014285601675510406, + 0.008938916027545929, + 0.007752656936645508, + 0.006225821562111378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.005254624877125025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, low-angle shot, rendered with photorealistic CGI cinematic realism, slowly advances onto the surface of a violent, roiling alien sea, dominated by a massive, angular Hunter-Warrior starship. Mountains of viscous, inky-black water undulate, their crests exploding into chaotic foam beneath a sky choked by scarred, dark gray storm clouds intermittently illuminated by silent, branching bursts of purple lightning. The starship's black hull, a masterpiece of hostile geometry, constructed from a matte, non-reflective composite, relentlessly slices through towering waves, sending volcanic spray blasting into the air. Dramatic lighting cast by strobe lightning captures the atomized spray and foam, causing them to shimmer like shattered glass against the vessel's obsidian form and create a stark, shifting black-and-white contrast effect, momentarily revealing the intricate texture of its armor plating. The ship's immense scale is palpable, its unstoppable forward momentum contrasting sharply with the raw, chaotic force of the ocean, creating an atmosphere of tension and awe-inspiring, overwhelming power.", + "seed": 11285, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_6/index_1651_seed_11285_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16085677491500974, + "vlm_logprob": -0.16085677491500974, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.16085677491500974, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10269346833229065 + ], + "opposite_logprob_tokens": [ + -2.727693557739258 + ], + "logprob_reward": -0.10269346833229065, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.10269346833229065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14218930900096893, + 0.05726296454668045, + 0.08188600838184357, + 0.05230489373207092, + 0.049631718546152115, + 0.1874588429927826, + 0.06870490312576294, + 0.24531276524066925, + 0.04938808083534241, + 0.0658605545759201 + ], + "credit_map": [ + 0.008861206471920013, + 0.015299760736525059, + 0.022710151970386505, + 0.017917219549417496, + 0.01785190775990486, + 0.017804155126214027, + 0.02391807921230793, + 0.01782681606709957, + 0.0034442199394106865, + 0.009818029589951038, + 0.005098063964396715, + 0.008101031184196472, + 0.008808927610516548, + 0.0061187176033854485, + 0.008269836194813251, + 0.007604140322655439, + 0.009467522613704205, + 0.010888087563216686, + 0.010550621896982193, + 0.008529800921678543, + 0.010973733849823475, + 0.00927218608558178, + 0.011739310808479786, + 0.010464748367667198, + 0.005908966530114412, + 0.006722075864672661, + 0.008134986273944378, + 0.006825404241681099, + 0.006185214035212994, + 0.006365671753883362, + 0.005032655317336321, + 0.007129918318241835, + 0.005317716859281063, + 0.009345653466880322, + 0.00452687032520771, + 0.006612156517803669, + 0.008334034122526646, + 0.006565234623849392, + 0.003546855179592967, + 0.005383199080824852, + 0.029252801090478897, + 0.02717732824385166, + 0.012436944991350174, + 0.01419716514647007, + 0.029819341376423836, + 0.03407169505953789, + 0.020078275352716446, + 0.020425286144018173, + 0.0074802725575864315, + 0.012551638297736645, + 0.007618834730237722, + 0.007103279232978821, + 0.012510704807937145, + 0.01162837352603674, + 0.004249506164342165, + 0.00556229380890727, + 0.011622767895460129, + 0.01151757501065731, + 0.00393245043233037, + 0.014276434667408466, + 0.09996389597654343, + 0.09351121634244919, + 0.0035234580282121897, + 0.006964969914406538, + 0.01095676701515913, + 0.00719913886860013, + 0.005525801330804825, + 0.004530760925263166, + 0.0090227872133255, + 0.005674137268215418, + 0.00437950249761343, + 0.0020991836208850145, + 0.013332645408809185, + 0.008702024817466736, + 0.0032597051467746496, + 0.0036820564419031143, + 0.00885984767228365, + 0.0072341859340667725, + 0.012158367782831192, + 0.008631719276309013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.10269346833229065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16044963896274567 + ], + "opposite_logprob_tokens": [ + -1.9104496240615845 + ], + "logprob_reward": -0.16044963896274567, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16044963896274567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1446964144706726, + 0.04069720208644867, + 0.07601537555456161, + 0.03298497200012207, + 0.026852430775761604, + 0.09391581267118454, + 0.03945843502879143, + 0.4018719792366028, + 0.06814970076084137, + 0.07535772025585175 + ], + "credit_map": [ + 0.008614836260676384, + 0.015853745862841606, + 0.020558763295412064, + 0.012972299940884113, + 0.022632718086242676, + 0.020090222358703613, + 0.028380878269672394, + 0.015592954121530056, + 0.0032804289367049932, + 0.00439772242680192, + 0.005440392531454563, + 0.006838830187916756, + 0.00558563694357872, + 0.005092841573059559, + 0.0057126665487885475, + 0.0043486859649419785, + 0.005420254077762365, + 0.007559752091765404, + 0.01220909133553505, + 0.013702522031962872, + 0.009722956456243992, + 0.008731470443308353, + 0.008593263104557991, + 0.01007606741040945, + 0.003397570690140128, + 0.00358506734482944, + 0.00499989977106452, + 0.00622602691873908, + 0.0036318027414381504, + 0.004388772416859865, + 0.0034348112531006336, + 0.003321017837151885, + 0.003384572919458151, + 0.004015249200165272, + 0.0023274943232536316, + 0.003685693722218275, + 0.003008082741871476, + 0.0032534922938793898, + 0.0030715514440089464, + 0.004106294829398394, + 0.009253110736608505, + 0.01197765488177538, + 0.00940652284771204, + 0.010671467520296574, + 0.01508273370563984, + 0.01485616248100996, + 0.013195398263633251, + 0.009472758509218693, + 0.005528334993869066, + 0.005644964519888163, + 0.0038602743297815323, + 0.00396911334246397, + 0.006907972041517496, + 0.007099600974470377, + 0.0025341780856251717, + 0.003913995809853077, + 0.022089403122663498, + 0.020356029272079468, + 0.011863668449223042, + 0.024648243561387062, + 0.1593800187110901, + 0.13994885981082916, + 0.009002096951007843, + 0.014583628624677658, + 0.015199597924947739, + 0.008193597197532654, + 0.008546131663024426, + 0.006763593293726444, + 0.010635480284690857, + 0.006362216547131538, + 0.008370818570256233, + 0.004078255966305733, + 0.014066088013350964, + 0.009052954614162445, + 0.004792219493538141, + 0.004945855122059584, + 0.010540763847529888, + 0.007928842678666115, + 0.014052597805857658, + 0.009978394955396652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.16044963896274567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256047938019037 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256047938019037, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256047938019037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11458783596754074, + 0.07395099103450775, + 0.06681807339191437, + 0.04010468348860741, + 0.12638205289840698, + 0.08794893324375153, + 0.14089316129684448, + 0.2348499596118927, + 0.039716511964797974, + 0.07474783807992935 + ], + "credit_map": [ + 0.006199791096150875, + 0.009907880797982216, + 0.015632998198270798, + 0.01655477099120617, + 0.01365853101015091, + 0.017817001789808273, + 0.02175269089639187, + 0.013064175844192505, + 0.006640241947025061, + 0.011353283189237118, + 0.005779585335403681, + 0.011160017922520638, + 0.015033661387860775, + 0.009635087102651596, + 0.008075741119682789, + 0.006273378152400255, + 0.006360115949064493, + 0.008038836531341076, + 0.009483888745307922, + 0.005314695183187723, + 0.0076815783977508545, + 0.008502393029630184, + 0.010790211148560047, + 0.010646358132362366, + 0.004772492218762636, + 0.0050008599646389484, + 0.0050004213117063046, + 0.004068592097610235, + 0.006328526884317398, + 0.004998289979994297, + 0.004905508365482092, + 0.005029991269111633, + 0.013518772087991238, + 0.024587661027908325, + 0.00781551655381918, + 0.010916103608906269, + 0.02042035199701786, + 0.018153144046664238, + 0.009584479033946991, + 0.021386027336120605, + 0.009170366451144218, + 0.012664029374718666, + 0.005973347928375006, + 0.0073956879787147045, + 0.01598977856338024, + 0.013140509836375713, + 0.013790157623589039, + 0.009825046174228191, + 0.017795301973819733, + 0.019591201096773148, + 0.015177702531218529, + 0.018689848482608795, + 0.022229166701436043, + 0.023209068924188614, + 0.011215077713131905, + 0.01298578456044197, + 0.014405694790184498, + 0.018315181136131287, + 0.011856218799948692, + 0.01646510325372219, + 0.08412452787160873, + 0.07444146275520325, + 0.006770576816052198, + 0.008471188135445118, + 0.00609250832349062, + 0.005068936385214329, + 0.006033057346940041, + 0.005638508591800928, + 0.00628636684268713, + 0.00436236709356308, + 0.004030784126371145, + 0.0022039837203919888, + 0.016952333971858025, + 0.009368433617055416, + 0.0031956578604876995, + 0.003227807115763426, + 0.010412781499326229, + 0.007268588058650494, + 0.014263071119785309, + 0.010059167630970478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.005256047938019037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233463048934937 + ], + "opposite_logprob_tokens": [ + -0.8983463048934937 + ], + "logprob_reward": -0.5233463048934937, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233463048934937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17036324739456177, + 0.09092317521572113, + 0.03350925073027611, + 0.044318147003650665, + 0.05194225162267685, + 0.03748338297009468, + 0.11686106026172638, + 0.2835431396961212, + 0.07931256294250488, + 0.09174377471208572 + ], + "credit_map": [ + 0.01037906389683485, + 0.01703369989991188, + 0.028291873633861542, + 0.02010309509932995, + 0.016060134395956993, + 0.031059592962265015, + 0.023086005821824074, + 0.024349769577383995, + 0.009337713941931725, + 0.01039209682494402, + 0.011168505065143108, + 0.017488814890384674, + 0.01194979902356863, + 0.010720610618591309, + 0.011254173703491688, + 0.008611458353698254, + 0.00325757940299809, + 0.003872237168252468, + 0.005944091826677322, + 0.003289317712187767, + 0.0037014512345194817, + 0.003880790900439024, + 0.00425318069756031, + 0.0053106024861335754, + 0.005751859396696091, + 0.006326009519398212, + 0.007688107434660196, + 0.004744342528283596, + 0.005753516685217619, + 0.006018490064889193, + 0.0036271915305405855, + 0.004408632405102253, + 0.006711642723530531, + 0.008930525742471218, + 0.0031767534092068672, + 0.0055250246077775955, + 0.009137243963778019, + 0.0077235582284629345, + 0.004981574602425098, + 0.005755929742008448, + 0.005463352892547846, + 0.004519242327660322, + 0.003016896080225706, + 0.003908992279320955, + 0.006102238781750202, + 0.006694705691188574, + 0.0037588144186884165, + 0.00401914119720459, + 0.010332931764423847, + 0.016664117574691772, + 0.008543383330106735, + 0.017101500183343887, + 0.018335936591029167, + 0.017422307282686234, + 0.008989213034510612, + 0.0194716714322567, + 0.01626739837229252, + 0.017023464664816856, + 0.012755271978676319, + 0.016991663724184036, + 0.10737413913011551, + 0.09696787595748901, + 0.00606605876237154, + 0.010097263380885124, + 0.012263139709830284, + 0.009600815363228321, + 0.011757735162973404, + 0.010293898172676563, + 0.013927393592894077, + 0.009880810976028442, + 0.00746507290750742, + 0.004123697057366371, + 0.01729900948703289, + 0.012153133749961853, + 0.0056300838477909565, + 0.0053496346808969975, + 0.012869519181549549, + 0.01207607239484787, + 0.013845633715391159, + 0.012520684860646725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.5233463048934937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253841444849968 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.01253841444849968, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253841444849968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10808002948760986, + 0.14023244380950928, + 0.29701048135757446, + 0.09534958750009537, + 0.027615604922175407, + 0.055752430111169815, + 0.05557764321565628, + 0.06054462492465973, + 0.051064811646938324, + 0.10877226293087006 + ], + "credit_map": [ + 0.009590775705873966, + 0.015039985999464989, + 0.013845482841134071, + 0.01599261350929737, + 0.01103392243385315, + 0.013483048416674137, + 0.015010195784270763, + 0.014084012247622013, + 0.012407653965055943, + 0.01827644743025303, + 0.015576374717056751, + 0.01588134467601776, + 0.028514297679066658, + 0.01784922555088997, + 0.01823555864393711, + 0.013491552323102951, + 0.0301669891923666, + 0.040035005658864975, + 0.03208579495549202, + 0.025732554495334625, + 0.04078114777803421, + 0.03480732440948486, + 0.059425435960292816, + 0.03397626057267189, + 0.012231886386871338, + 0.009432826191186905, + 0.015581563115119934, + 0.014151922427117825, + 0.013067660853266716, + 0.013613459654152393, + 0.0062508294358849525, + 0.011019456200301647, + 0.00408039428293705, + 0.00448734313249588, + 0.002316829049959779, + 0.0030096450354903936, + 0.003990830387920141, + 0.003847840940579772, + 0.0024708283599466085, + 0.003411895362660289, + 0.005235396325588226, + 0.007938379421830177, + 0.003596898401156068, + 0.004796513821929693, + 0.011889897286891937, + 0.00854578148573637, + 0.007600195240229368, + 0.006149365566670895, + 0.007236456964164972, + 0.00781677570194006, + 0.004983701277524233, + 0.005259130150079727, + 0.009561066515743732, + 0.0099523039534688, + 0.003826118540018797, + 0.00694209523499012, + 0.00698184035718441, + 0.008466795086860657, + 0.006143258884549141, + 0.005833709612488747, + 0.012820500880479813, + 0.011198543943464756, + 0.0038194519001990557, + 0.00528053380548954, + 0.00814354233443737, + 0.010035048238933086, + 0.004610130563378334, + 0.006141500547528267, + 0.00918859988451004, + 0.006559708621352911, + 0.0037476541474461555, + 0.0026386333629488945, + 0.030964650213718414, + 0.014618678018450737, + 0.005586082581430674, + 0.0059919338673353195, + 0.012447639368474483, + 0.00896734744310379, + 0.020868105813860893, + 0.009327836334705353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.01253841444849968 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, low-angle shot, rendered with photorealistic CGI cinematic realism, slowly advances onto the surface of a violent, roiling alien sea, dominated by a massive, angular Hunter-Warrior starship. Mountains of viscous, inky-black water undulate, their crests exploding into chaotic foam beneath a sky choked by scarred, dark gray storm clouds intermittently illuminated by silent, branching bursts of purple lightning. The starship's black hull, a masterpiece of hostile geometry, constructed from a matte, non-reflective composite, relentlessly slices through towering waves, sending volcanic spray blasting into the air. Dramatic lighting cast by strobe lightning captures the atomized spray and foam, causing them to shimmer like shattered glass against the vessel's obsidian form and create a stark, shifting black-and-white contrast effect, momentarily revealing the intricate texture of its armor plating. The ship's immense scale is palpable, its unstoppable forward momentum contrasting sharply with the raw, chaotic force of the ocean, creating an atmosphere of tension and awe-inspiring, overwhelming power.", + "seed": 11286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_6/index_1651_seed_11286_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09848095010966063, + "vlm_logprob": -0.09848095010966063, + "vlm_yes_no_margin": 3.150000011920929, + "vlm_token_credit": -0.09848095010966063, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03870507702231407 + ], + "opposite_logprob_tokens": [ + -3.9137051105499268 + ], + "logprob_reward": -0.03870507702231407, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03870507702231407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06968844681978226, + 0.11708926409482956, + 0.47275131940841675, + 0.03944237902760506, + 0.043795276433229446, + 0.03440143167972565, + 0.15628419816493988, + 0.023319460451602936, + 0.023930290713906288, + 0.019297853112220764 + ], + "credit_map": [ + 0.004309547133743763, + 0.004785298369824886, + 0.010859841480851173, + 0.012259647250175476, + 0.007936527021229267, + 0.007403214927762747, + 0.011266603134572506, + 0.010867773555219173, + 0.015331421047449112, + 0.02756795659661293, + 0.009864097461104393, + 0.009170493111014366, + 0.015871679410338402, + 0.013756605796515942, + 0.013641230762004852, + 0.011885790154337883, + 0.029458511620759964, + 0.015801552683115005, + 0.014014670625329018, + 0.017296934500336647, + 0.011528578586876392, + 0.014575953595340252, + 0.16968883574008942, + 0.20038631558418274, + 0.00279309693723917, + 0.0022914966102689505, + 0.006108707282692194, + 0.0031946541275829077, + 0.003563803154975176, + 0.004344669170677662, + 0.0072385817766189575, + 0.009907371364533901, + 0.0032535402569919825, + 0.0025179164949804544, + 0.0040129926055669785, + 0.005862507037818432, + 0.003835007082670927, + 0.004757254384458065, + 0.01129261963069439, + 0.00826344545930624, + 0.003977422136813402, + 0.003801684593781829, + 0.003464391455054283, + 0.005667627789080143, + 0.004969342611730099, + 0.0031651128083467484, + 0.004433590453118086, + 0.0049222661182284355, + 0.011996145360171795, + 0.012396568432450294, + 0.05065200477838516, + 0.014129316434264183, + 0.014419155195355415, + 0.0076436190865933895, + 0.008406735956668854, + 0.03664066642522812, + 0.0019288032781332731, + 0.0029162191785871983, + 0.0038149601314216852, + 0.0027970331721007824, + 0.0035997838713228703, + 0.00376066192984581, + 0.002550693927332759, + 0.0019513070583343506, + 0.0015396429225802422, + 0.00198066933080554, + 0.0015828657196834683, + 0.007234420161694288, + 0.0015310740564018488, + 0.0016115527832880616, + 0.00654135225340724, + 0.00190871546510607, + 0.0018026820616796613, + 0.003114613937214017, + 0.0022651641629636288, + 0.0029061834793537855, + 0.003243869636207819, + 0.0027062054723501205, + 0.0017359612975269556, + 0.0015231730649247766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.03870507702231407 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005749642848969 + ], + "opposite_logprob_tokens": [ + -2.6950573921203613 + ], + "logprob_reward": -0.07005749642848969, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005749642848969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060226310044527054, + 0.05224720388650894, + 0.26398736238479614, + 0.08475358784198761, + 0.025607574731111526, + 0.06022452563047409, + 0.307333767414093, + 0.03871260583400726, + 0.07465925067663193, + 0.03224780410528183 + ], + "credit_map": [ + 0.005136453080922365, + 0.008533470332622528, + 0.005604530684649944, + 0.006110336631536484, + 0.01245757658034563, + 0.007367085665464401, + 0.006890700198709965, + 0.008126159198582172, + 0.004287556279450655, + 0.008844884112477303, + 0.0053376685827970505, + 0.006013125646859407, + 0.006421377882361412, + 0.009229983203113079, + 0.006135862786322832, + 0.0059767430648207664, + 0.014349090866744518, + 0.008603505790233612, + 0.013102992437779903, + 0.01340540312230587, + 0.008012201637029648, + 0.008979536592960358, + 0.0895046666264534, + 0.10802997648715973, + 0.005817900877445936, + 0.0036194680724292994, + 0.016993239521980286, + 0.007847961038351059, + 0.006390250287950039, + 0.006939080078154802, + 0.015420963987708092, + 0.021724723279476166, + 0.00205785664729774, + 0.0021530676167458296, + 0.002690378110855818, + 0.003463699482381344, + 0.003989534918218851, + 0.002966729225590825, + 0.005052118096500635, + 0.0032341915648430586, + 0.006600941531360149, + 0.007280598860234022, + 0.005773074924945831, + 0.007476428523659706, + 0.008851919323205948, + 0.006415807642042637, + 0.009637100622057915, + 0.008188655599951744, + 0.024340258911252022, + 0.01664155349135399, + 0.11156462132930756, + 0.024355754256248474, + 0.020578252151608467, + 0.015077575109899044, + 0.01510606985539198, + 0.07966967672109604, + 0.0036800650414079428, + 0.005731855519115925, + 0.006429281551390886, + 0.00542966416105628, + 0.006212278734892607, + 0.004112056456506252, + 0.0040076603181660175, + 0.0031097428873181343, + 0.006058062426745892, + 0.006970848888158798, + 0.00509229302406311, + 0.022150753065943718, + 0.010364734567701817, + 0.007127857301384211, + 0.012069525197148323, + 0.004825182259082794, + 0.003328882623463869, + 0.004085900261998177, + 0.004236339591443539, + 0.003231816226616502, + 0.004265674855560064, + 0.006627507507801056, + 0.003320490475744009, + 0.003151193493977189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.07005749642848969 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344336569309235 + ], + "opposite_logprob_tokens": [ + -3.6513442993164062 + ], + "logprob_reward": -0.026344336569309235, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344336569309235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05489242449402809, + 0.031879838556051254, + 0.2603866755962372, + 0.13203726708889008, + 0.025634942576289177, + 0.03443579375743866, + 0.06399405002593994, + 0.22027088701725006, + 0.16232068836688995, + 0.01414745394140482 + ], + "credit_map": [ + 0.004865891300141811, + 0.005074723623692989, + 0.003934324719011784, + 0.0051170471124351025, + 0.013743679970502853, + 0.00858455803245306, + 0.007086608558893204, + 0.006485592108219862, + 0.0028166461270302534, + 0.00473876902833581, + 0.0033260569907724857, + 0.003999419044703245, + 0.0035616906825453043, + 0.005085111130028963, + 0.003973411861807108, + 0.004378731828182936, + 0.016631802543997765, + 0.010369419120252132, + 0.010202275589108467, + 0.010404837317764759, + 0.012822671793401241, + 0.013772087171673775, + 0.0859576016664505, + 0.10022597759962082, + 0.011301778256893158, + 0.006222270894795656, + 0.02897913195192814, + 0.012248883955180645, + 0.007686140947043896, + 0.006754585076123476, + 0.024001045152544975, + 0.034843429923057556, + 0.0018527043284848332, + 0.001431895885616541, + 0.003089746693149209, + 0.0030344652477651834, + 0.0024069086648523808, + 0.002769039012491703, + 0.0064780679531395435, + 0.004572116304188967, + 0.0028124814853072166, + 0.0038942433893680573, + 0.0027068788185715675, + 0.005274519789963961, + 0.0053601348772645, + 0.0027693570591509342, + 0.006218193098902702, + 0.005399984307587147, + 0.004995308816432953, + 0.005415491759777069, + 0.013790945522487164, + 0.009347287937998772, + 0.009934510104358196, + 0.004382539074867964, + 0.005467529874294996, + 0.010660435073077679, + 0.017553506419062614, + 0.02851094864308834, + 0.03319545462727547, + 0.028024306520819664, + 0.035115037113428116, + 0.039169564843177795, + 0.020680105313658714, + 0.018021956086158752, + 0.00816312525421381, + 0.016415078192949295, + 0.016108907759189606, + 0.03149119392037392, + 0.017009271308779716, + 0.025041190907359123, + 0.035749029368162155, + 0.012342900037765503, + 0.0012527492363005877, + 0.0019244455033913255, + 0.0016765289474278688, + 0.001648991135880351, + 0.0015369971515610814, + 0.0029961029067635536, + 0.0016180165112018585, + 0.0014936216175556183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.026344336569309235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486582636833191 + ], + "opposite_logprob_tokens": [ + -1.2236583232879639 + ], + "logprob_reward": -0.3486582636833191, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3486582636833191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11978595703840256, + 0.09602756053209305, + 0.14084599912166595, + 0.1465405374765396, + 0.10048546642065048, + 0.024248726665973663, + 0.0921449363231659, + 0.0791373997926712, + 0.16941073536872864, + 0.03137272596359253 + ], + "credit_map": [ + 0.00862820167094469, + 0.014123420231044292, + 0.012505656108260155, + 0.014093234203755856, + 0.02623450756072998, + 0.01199919730424881, + 0.01709086075425148, + 0.015110875479876995, + 0.011304649524390697, + 0.022711656987667084, + 0.008149223402142525, + 0.007136307191103697, + 0.014035089872777462, + 0.011264216154813766, + 0.011874737218022346, + 0.00955168530344963, + 0.008330375887453556, + 0.007111710496246815, + 0.008292985148727894, + 0.007391592022031546, + 0.014439906924962997, + 0.01900893822312355, + 0.035041142255067825, + 0.04122934490442276, + 0.017563769593834877, + 0.008348285220563412, + 0.03947281837463379, + 0.0175658967345953, + 0.010113297030329704, + 0.011439821682870388, + 0.01600565016269684, + 0.026031002402305603, + 0.010360503569245338, + 0.004967250395566225, + 0.006465437822043896, + 0.010402414947748184, + 0.004731297958642244, + 0.005538477096706629, + 0.027174662798643112, + 0.03084542788565159, + 0.0024242575746029615, + 0.0019474962027743459, + 0.0022287701722234488, + 0.0033336926717311144, + 0.004163575358688831, + 0.0023460767697542906, + 0.004667914006859064, + 0.0031369440257549286, + 0.006854924373328686, + 0.005565274972468615, + 0.029945366084575653, + 0.008891102857887745, + 0.007939382456243038, + 0.005954557564109564, + 0.005325863603502512, + 0.02166847325861454, + 0.007848285138607025, + 0.012251741252839565, + 0.0082775317132473, + 0.010042145848274231, + 0.012153474614024162, + 0.014270228333771229, + 0.008558603003621101, + 0.0057353968732059, + 0.007952055893838406, + 0.011756359599530697, + 0.010289023630321026, + 0.044660747051239014, + 0.020294342190027237, + 0.012283158488571644, + 0.0524214431643486, + 0.009753613732755184, + 0.0031573064625263214, + 0.005085871554911137, + 0.004417938180267811, + 0.0035009540151804686, + 0.003609880106523633, + 0.004397459328174591, + 0.004298292100429535, + 0.0029050270095467567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.3486582636833191 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639576844871044 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639576844871044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639576844871044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1419285237789154, + 0.09028542041778564, + 0.10486579686403275, + 0.3031545579433441, + 0.076774001121521, + 0.039234988391399384, + 0.10319580137729645, + 0.03970417007803917, + 0.07098840177059174, + 0.02986840531229973 + ], + "credit_map": [ + 0.01278845313936472, + 0.013892889954149723, + 0.012912517413496971, + 0.017466936260461807, + 0.029694603756070137, + 0.01583469659090042, + 0.016912570223212242, + 0.02242586389183998, + 0.011167054995894432, + 0.019217388704419136, + 0.009956813417375088, + 0.007180484477430582, + 0.011138459667563438, + 0.012483377009630203, + 0.01063576340675354, + 0.008506085723638535, + 0.006406425032764673, + 0.005222488660365343, + 0.003960329107940197, + 0.003922910429537296, + 0.006722571793943644, + 0.007968803867697716, + 0.03223053738474846, + 0.03843173012137413, + 0.01747451350092888, + 0.012522912584245205, + 0.04211617633700371, + 0.02210838347673416, + 0.02310493215918541, + 0.03912876546382904, + 0.06452755630016327, + 0.08217128366231918, + 0.008370892144739628, + 0.005191314034163952, + 0.007337110582739115, + 0.011585195548832417, + 0.007778731640428305, + 0.008908803574740887, + 0.015752023085951805, + 0.011849932372570038, + 0.004121402744203806, + 0.005796595010906458, + 0.003954187035560608, + 0.005025704391300678, + 0.005581437144428492, + 0.004187646321952343, + 0.004744344390928745, + 0.005823672283440828, + 0.008555203676223755, + 0.0062356679700315, + 0.032581210136413574, + 0.008511696010828018, + 0.007402036804705858, + 0.007772302255034447, + 0.006822558119893074, + 0.02531513012945652, + 0.0034120541531592607, + 0.00632808543741703, + 0.006400115787982941, + 0.005081556737422943, + 0.005628213286399841, + 0.004449031315743923, + 0.004901280160993338, + 0.003503836691379547, + 0.007225247099995613, + 0.007891140878200531, + 0.005658573471009731, + 0.009889677166938782, + 0.015257606282830238, + 0.011661331169307232, + 0.006071343086659908, + 0.007333475165069103, + 0.0024738262873142958, + 0.004061568528413773, + 0.004107266664505005, + 0.004428540822118521, + 0.004013575613498688, + 0.004642011132091284, + 0.0029129418544471264, + 0.0032286716159433126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.008639576844871044 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, low-angle shot, rendered with photorealistic CGI cinematic realism, slowly advances onto the surface of a violent, roiling alien sea, dominated by a massive, angular Hunter-Warrior starship. Mountains of viscous, inky-black water undulate, their crests exploding into chaotic foam beneath a sky choked by scarred, dark gray storm clouds intermittently illuminated by silent, branching bursts of purple lightning. The starship's black hull, a masterpiece of hostile geometry, constructed from a matte, non-reflective composite, relentlessly slices through towering waves, sending volcanic spray blasting into the air. Dramatic lighting cast by strobe lightning captures the atomized spray and foam, causing them to shimmer like shattered glass against the vessel's obsidian form and create a stark, shifting black-and-white contrast effect, momentarily revealing the intricate texture of its armor plating. The ship's immense scale is palpable, its unstoppable forward momentum contrasting sharply with the raw, chaotic force of the ocean, creating an atmosphere of tension and awe-inspiring, overwhelming power.", + "seed": 11287, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_6/index_1651_seed_11287_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13503125738352537, + "vlm_logprob": -0.13503125738352537, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.13503125738352537, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058398064225912094 + ], + "opposite_logprob_tokens": [ + -3.3083980083465576 + ], + "logprob_reward": -0.058398064225912094, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.058398064225912094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5873359441757202, + 0.02657281421124935, + 0.05234922468662262, + 0.04651101306080818, + 0.030184369534254074, + 0.03798188269138336, + 0.033886827528476715, + 0.04228765890002251, + 0.06422899663448334, + 0.0786612406373024 + ], + "credit_map": [ + 0.0603192001581192, + 0.08426988124847412, + 0.02748080901801586, + 0.04376325011253357, + 0.15596790611743927, + 0.08621452748775482, + 0.06364860385656357, + 0.06567173451185226, + 0.0035298962611705065, + 0.005027457140386105, + 0.002732094842940569, + 0.002765798242762685, + 0.004025241360068321, + 0.0030888875480741262, + 0.002929005539044738, + 0.002474432112649083, + 0.00433378154411912, + 0.0042269788682460785, + 0.0048056417144834995, + 0.004162991885095835, + 0.009762403555214405, + 0.014690736308693886, + 0.005162630695849657, + 0.0052040573209524155, + 0.005077261012047529, + 0.00430967565625906, + 0.004168176557868719, + 0.006987585220485926, + 0.005409098695963621, + 0.00989806279540062, + 0.007371065206825733, + 0.0032900897786021233, + 0.00361440796405077, + 0.005148111376911402, + 0.002396122319623828, + 0.002900197170674801, + 0.004441119264811277, + 0.0049186693504452705, + 0.003769987029954791, + 0.00299575156532228, + 0.0032219584099948406, + 0.004272888880223036, + 0.0023417260963469744, + 0.0059685418382287025, + 0.00575457327067852, + 0.004277708008885384, + 0.008054831996560097, + 0.004089654423296452, + 0.002875403268262744, + 0.0046760872937738895, + 0.0020313234999775887, + 0.003891794942319393, + 0.005612354725599289, + 0.007535209879279137, + 0.004116493742913008, + 0.003148159245029092, + 0.0045868586748838425, + 0.0054915063083171844, + 0.006676181685179472, + 0.005513045936822891, + 0.005284878425300121, + 0.0046559665352106094, + 0.0057349312119185925, + 0.004344290588051081, + 0.01042984426021576, + 0.007146660704165697, + 0.007678098510950804, + 0.006973963230848312, + 0.007373409811407328, + 0.007582432124763727, + 0.008240127004683018, + 0.008804455399513245, + 0.011617007665336132, + 0.012928894720971584, + 0.007095579523593187, + 0.008220594376325607, + 0.01015656441450119, + 0.008873173967003822, + 0.011059693992137909, + 0.008709732443094254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.058398064225912094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903339713811874 + ], + "opposite_logprob_tokens": [ + -2.579033374786377 + ], + "logprob_reward": -0.07903339713811874, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903339713811874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1558341085910797, + 0.052519869059324265, + 0.09126391261816025, + 0.08699686825275421, + 0.04333767294883728, + 0.04347244277596474, + 0.06185637786984444, + 0.19445163011550903, + 0.15249280631542206, + 0.11777429282665253 + ], + "credit_map": [ + 0.01651998795568943, + 0.026125473901629448, + 0.01227122824639082, + 0.01261430699378252, + 0.03953113779425621, + 0.022121388465166092, + 0.013524125330150127, + 0.013126472011208534, + 0.006424340885132551, + 0.009397557936608791, + 0.00808016862720251, + 0.00634951563552022, + 0.004930844996124506, + 0.006865243893116713, + 0.0054026260040700436, + 0.00506957434117794, + 0.00927689578384161, + 0.012969120405614376, + 0.008587085641920567, + 0.008667133748531342, + 0.013729263097047806, + 0.016996903344988823, + 0.012035082094371319, + 0.009002435952425003, + 0.009279487654566765, + 0.009878703393042088, + 0.006417163647711277, + 0.010737292468547821, + 0.010201248340308666, + 0.02151075005531311, + 0.012338432483375072, + 0.006633801385760307, + 0.0044270590879023075, + 0.005157419014722109, + 0.0029335850849747658, + 0.005205526016652584, + 0.007010666187852621, + 0.006679770536720753, + 0.007666097953915596, + 0.004257551860064268, + 0.004208737518638372, + 0.004803590476512909, + 0.002625850262120366, + 0.005192648619413376, + 0.00466802017763257, + 0.005944621749222279, + 0.011869954876601696, + 0.004159025847911835, + 0.011183985508978367, + 0.005274858325719833, + 0.008883632719516754, + 0.00810297578573227, + 0.006136552430689335, + 0.008741158060729504, + 0.005637691356241703, + 0.007895532064139843, + 0.01629631035029888, + 0.02274378016591072, + 0.0168233010917902, + 0.0412677563726902, + 0.020917942747473717, + 0.018658632412552834, + 0.03970402479171753, + 0.018039902672171593, + 0.021621573716402054, + 0.015240729786455631, + 0.012231683358550072, + 0.014919155277311802, + 0.027823997661471367, + 0.03408294916152954, + 0.012870210222899914, + 0.0137025136500597, + 0.013719545677304268, + 0.015260954387485981, + 0.00929808896034956, + 0.012489738874137402, + 0.014408513903617859, + 0.025590643286705017, + 0.01620149053633213, + 0.010805334895849228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.07903339713811874 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06681716442108154, + 0.018873993307352066, + 0.04182460904121399, + 0.09718553721904755, + 0.01726066879928112, + 0.03090544603765011, + 0.020087221637368202, + 0.38022586703300476, + 0.07211699336767197, + 0.25470250844955444 + ], + "credit_map": [ + 0.0072890338487923145, + 0.008961520157754421, + 0.004225437995046377, + 0.005014414899051189, + 0.017492100596427917, + 0.009992319159209728, + 0.006669648922979832, + 0.007172686979174614, + 0.0029181595891714096, + 0.004107182379812002, + 0.0016409100499004126, + 0.001613994361832738, + 0.0018447801703587174, + 0.002702706726267934, + 0.0025476582814007998, + 0.001498599536716938, + 0.002784600481390953, + 0.004487946163862944, + 0.0027934915851801634, + 0.002815505489706993, + 0.011124324053525925, + 0.010869337245821953, + 0.003240377176553011, + 0.003709028474986553, + 0.01056445948779583, + 0.009635364636778831, + 0.0069683450274169445, + 0.011853023432195187, + 0.01089564710855484, + 0.02697829157114029, + 0.012636950239539146, + 0.007653449662029743, + 0.002680232748389244, + 0.0021767173893749714, + 0.001395837520249188, + 0.0019060930935665965, + 0.002236882923170924, + 0.002739206189289689, + 0.0021855856757611036, + 0.0019401125609874725, + 0.0021908392664045095, + 0.0037684522103518248, + 0.0018067682394757867, + 0.004208751488476992, + 0.006003139074891806, + 0.004389258101582527, + 0.005434685852378607, + 0.003103553084656596, + 0.0015838294057175517, + 0.0022045730147510767, + 0.0015865301247686148, + 0.0034664850682020187, + 0.0027608503587543964, + 0.00420094421133399, + 0.002021999564021826, + 0.002262010471895337, + 0.04093456268310547, + 0.04814745858311653, + 0.0361676849424839, + 0.0783994197845459, + 0.03586200997233391, + 0.035865116864442825, + 0.06825421750545502, + 0.036595407873392105, + 0.010213845409452915, + 0.007699030451476574, + 0.005468194838613272, + 0.0037294498179107904, + 0.01660589687526226, + 0.015140056610107422, + 0.006620642263442278, + 0.006639878731220961, + 0.02792670950293541, + 0.04370304197072983, + 0.019218115136027336, + 0.03502143174409866, + 0.036907926201820374, + 0.029178382828831673, + 0.038062673062086105, + 0.02468424290418625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.007631433662027121 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233557820320129 + ], + "opposite_logprob_tokens": [ + -0.8983557820320129 + ], + "logprob_reward": -0.5233557820320129, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233557820320129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1791231334209442, + 0.05865643545985222, + 0.07128733396530151, + 0.0620235800743103, + 0.05287601798772812, + 0.08184623718261719, + 0.043560173362493515, + 0.22793349623680115, + 0.061864469200372696, + 0.1608290821313858 + ], + "credit_map": [ + 0.015061997808516026, + 0.012419264763593674, + 0.01600147783756256, + 0.013529698364436626, + 0.041417695581912994, + 0.030179282650351524, + 0.02215312048792839, + 0.028360603377223015, + 0.007045028265565634, + 0.01516978070139885, + 0.006321968976408243, + 0.00637277914211154, + 0.005976776592433453, + 0.004895684774965048, + 0.008642563596367836, + 0.004231853410601616, + 0.007180328946560621, + 0.009723170660436153, + 0.010040502063930035, + 0.0059191337786614895, + 0.012127422727644444, + 0.01052853837609291, + 0.00842303317040205, + 0.007345201447606087, + 0.005827539134770632, + 0.012178714387118816, + 0.0059494758024811745, + 0.0068389358930289745, + 0.008465892635285854, + 0.010261059738695621, + 0.006977497134357691, + 0.005524462088942528, + 0.007935134693980217, + 0.00959507841616869, + 0.006396840792149305, + 0.005502280779182911, + 0.006994069088250399, + 0.006995085161179304, + 0.0047875358723104, + 0.004669989924877882, + 0.013390758074820042, + 0.010603073984384537, + 0.008522639982402325, + 0.00971271377056837, + 0.008745769038796425, + 0.013948353938758373, + 0.011255224235355854, + 0.005667705088853836, + 0.005591678898781538, + 0.005332421511411667, + 0.005513390991836786, + 0.005210311617702246, + 0.006516017485409975, + 0.005375784356147051, + 0.005985485389828682, + 0.004035080783069134, + 0.02084909938275814, + 0.029072506353259087, + 0.024456700310111046, + 0.046459317207336426, + 0.025521250441670418, + 0.02209942415356636, + 0.03741300106048584, + 0.022062193602323532, + 0.00995115377008915, + 0.009501275606453419, + 0.007406498305499554, + 0.004803817719221115, + 0.008744373917579651, + 0.0077475327998399734, + 0.007702491711825132, + 0.006007321644574404, + 0.019020631909370422, + 0.02285686321556568, + 0.010832230560481548, + 0.016548486426472664, + 0.02982254885137081, + 0.021905139088630676, + 0.020491275936365128, + 0.019351907074451447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.5233557820320129 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance onto the surface of a violent, roiling alien sea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive, angular Hunter-Warrior starship visible dominating the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the starship's hull black, matte, and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there scarred, dark gray storm clouds with intermittent purple lightning in the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673760985955596 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.00673760985955596, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673760985955596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30632683634757996, + 0.07813180238008499, + 0.030051255598664284, + 0.0775020495057106, + 0.02585504949092865, + 0.022528134286403656, + 0.09293150901794434, + 0.20334473252296448, + 0.08054772764444351, + 0.08278095722198486 + ], + "credit_map": [ + 0.031149445101618767, + 0.043475519865751266, + 0.02050000987946987, + 0.025515777990221977, + 0.062184784561395645, + 0.04627680778503418, + 0.03315749019384384, + 0.04406698793172836, + 0.009496081620454788, + 0.023237237706780434, + 0.006384897977113724, + 0.007995489053428173, + 0.009533402509987354, + 0.006129025481641293, + 0.009541425853967667, + 0.0058142333291471004, + 0.002826494164764881, + 0.003441327717155218, + 0.004371712915599346, + 0.002425161190330982, + 0.006174498237669468, + 0.005114795174449682, + 0.003473064163699746, + 0.0022241994738578796, + 0.008719473145902157, + 0.007390754763036966, + 0.006031987722963095, + 0.009764389134943485, + 0.008812900632619858, + 0.020939810201525688, + 0.010215581394731998, + 0.005627143196761608, + 0.003533534239977598, + 0.006038660183548927, + 0.0017977653769776225, + 0.002622688189148903, + 0.0032313342671841383, + 0.0036922511644661427, + 0.0025713155046105385, + 0.002367498818784952, + 0.0045037176460027695, + 0.003002413082867861, + 0.0019753454253077507, + 0.002333963755518198, + 0.003311960957944393, + 0.0030026871245354414, + 0.002825348172336817, + 0.001572697306983173, + 0.013631428591907024, + 0.011857590638101101, + 0.006262775044888258, + 0.006753674708306789, + 0.010721339844167233, + 0.025453632697463036, + 0.01022380217909813, + 0.00802726112306118, + 0.022310655564069748, + 0.025967013090848923, + 0.019319936633110046, + 0.041139695793390274, + 0.0205003023147583, + 0.01923980377614498, + 0.036624014377593994, + 0.018243303522467613, + 0.011538435705006123, + 0.009128307923674583, + 0.006041476037353277, + 0.006119220983237028, + 0.016824573278427124, + 0.018104312941432, + 0.007171614095568657, + 0.005619778763502836, + 0.009635193273425102, + 0.014640969224274158, + 0.006312721874564886, + 0.010679345577955246, + 0.011734076775610447, + 0.008073695935308933, + 0.012087441049516201, + 0.00961750652641058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.00673760985955596 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_tensor([3243, 3243])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_tensor([3243, 3243])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0b7b30aa2f29389a2c514160cdba9482c03ab011 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_tensor([3243, 3243])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in dramatic slow motion, this extreme close-up, tilting downward, follows a gleaming metal screw as it spirals through the air. The screw, a polished stainless steel Phillips head, tumbles gracefully, its meticulously machined threads catching a sharp, directional light from an unseen source above. This harsh light creates a mesmerizing display of brilliant, specular highlights that propagate along its metallic form as it rotates steadily about its longitudinal axis. The background is a dark, severely out-of-focus industrial space rendered in muted charcoal and dark brown, with motes of dust drifting lazily, catching the light and becoming soft, glowing bokeh balls. The dominant color palette is a high-contrast blend of the screw's cool, brilliant silver against the warm, dark void. The entire scene is imbued with a focused, almost poetic atmosphere, transforming the simple act of falling into a captivating ballet of light and movement into a captivating melody of light and movement.", + "seed": 10880, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_3243_seed_10880_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.033611895702779296, + "vlm_logprob": -0.033611895702779296, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.033611895702779296, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0396711640059948 + ], + "opposite_logprob_tokens": [ + -3.4146711826324463 + ], + "logprob_reward": -0.0396711640059948, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0396711640059948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12086854130029678, + 0.07466771453619003, + 0.10325198620557785, + 0.15606701374053955, + 0.06392182409763336, + 0.08098810911178589, + 0.0875815749168396, + 0.08864689618349075, + 0.13671503961086273, + 0.08729123324155807 + ], + "credit_map": [ + 0.012268269434571266, + 0.013113880529999733, + 0.009911535307765007, + 0.018678558990359306, + 0.017449723556637764, + 0.015128664672374725, + 0.02333945780992508, + 0.01097845658659935, + 0.006649385672062635, + 0.007478991523385048, + 0.01131061278283596, + 0.010850446298718452, + 0.009656007401645184, + 0.008311050944030285, + 0.009629277512431145, + 0.010781942866742611, + 0.012024985626339912, + 0.009118732064962387, + 0.020571254193782806, + 0.007570189889520407, + 0.009882357902824879, + 0.01113962009549141, + 0.0077886139042675495, + 0.025156227871775627, + 0.014856293797492981, + 0.00770858209580183, + 0.03982549533247948, + 0.015925725921988487, + 0.012574611231684685, + 0.008929552510380745, + 0.02155892737209797, + 0.0346878245472908, + 0.0056131393648684025, + 0.006640564650297165, + 0.009888501837849617, + 0.011577118188142776, + 0.004339565522968769, + 0.008122947998344898, + 0.007132770959287882, + 0.010607215575873852, + 0.008160139434039593, + 0.005373970605432987, + 0.016501786187291145, + 0.01151189673691988, + 0.004634992219507694, + 0.014034883119165897, + 0.009137124754488468, + 0.011633322574198246, + 0.011509047821164131, + 0.007318766787648201, + 0.012735961005091667, + 0.010445881634950638, + 0.008048677816987038, + 0.01229067798703909, + 0.0129933413118124, + 0.012239220552146435, + 0.008279634639620781, + 0.007793580647557974, + 0.016535313799977303, + 0.008477222174406052, + 0.007440667133778334, + 0.023245448246598244, + 0.007925573736429214, + 0.008949464187026024, + 0.009658053517341614, + 0.02067718654870987, + 0.03001929260790348, + 0.014896837063133717, + 0.007616785820573568, + 0.03634257614612579, + 0.007223941385746002, + 0.010280359536409378, + 0.006963703781366348, + 0.013168173842132092, + 0.014796075411140919, + 0.007336359471082687, + 0.01872928813099861, + 0.013825181871652603, + 0.0052030012011528015, + 0.007269449066370726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.0396711640059948 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026366282254457474 + ], + "opposite_logprob_tokens": [ + -3.6513662338256836 + ], + "logprob_reward": -0.026366282254457474, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366282254457474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11467505991458893, + 0.10459626466035843, + 0.07404681295156479, + 0.06341251730918884, + 0.04991523176431656, + 0.20667220652103424, + 0.060529906302690506, + 0.07316671311855316, + 0.16434447467327118, + 0.08864083141088486 + ], + "credit_map": [ + 0.011292539536952972, + 0.011123639531433582, + 0.010819358751177788, + 0.018917540088295937, + 0.014116368256509304, + 0.014583932235836983, + 0.02139151841402054, + 0.012430158443748951, + 0.009314258582890034, + 0.009559485130012035, + 0.011071844026446342, + 0.01327622402459383, + 0.02043760195374489, + 0.01291794702410698, + 0.01749003864824772, + 0.010528855957090855, + 0.0101247513666749, + 0.00954346265643835, + 0.011959216557443142, + 0.008374038152396679, + 0.007800477556884289, + 0.008815999142825603, + 0.006242286879569292, + 0.011186579242348671, + 0.007050969172269106, + 0.008306486532092094, + 0.009435383602976799, + 0.009808383882045746, + 0.0052114203572273254, + 0.010808294638991356, + 0.0061415741220116615, + 0.006650002673268318, + 0.004972377326339483, + 0.006387739907950163, + 0.008164341561496258, + 0.010874688625335693, + 0.003250816138461232, + 0.006490556988865137, + 0.004742003511637449, + 0.0050327046774327755, + 0.01726418174803257, + 0.013132181018590927, + 0.037687260657548904, + 0.027555353939533234, + 0.012714034877717495, + 0.03710377216339111, + 0.030832447111606598, + 0.030382953584194183, + 0.006532708182930946, + 0.005019377451390028, + 0.00658639008179307, + 0.008141271770000458, + 0.004160459619015455, + 0.009377075359225273, + 0.006330268457531929, + 0.014382353983819485, + 0.005640841089189053, + 0.0061413743533194065, + 0.017282653599977493, + 0.011600980535149574, + 0.006358318962156773, + 0.010543117299675941, + 0.007615574169903994, + 0.00798384752124548, + 0.011381505988538265, + 0.012052135542035103, + 0.0523056797683239, + 0.009139270521700382, + 0.008081549778580666, + 0.056078050285577774, + 0.005042864941060543, + 0.010263405740261078, + 0.005458254832774401, + 0.011782705783843994, + 0.0179962869733572, + 0.005626498255878687, + 0.02208980917930603, + 0.015589602291584015, + 0.004813509527593851, + 0.005284151528030634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.026366282254457474 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904840797185898 + ], + "opposite_logprob_tokens": [ + -2.464048385620117 + ], + "logprob_reward": -0.08904840797185898, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904840797185898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06423722952604294, + 0.05878009274601936, + 0.10670080780982971, + 0.08543679118156433, + 0.0625501498579979, + 0.3841795027256012, + 0.1388404220342636, + 0.03469393402338028, + 0.01408751867711544, + 0.050493523478507996 + ], + "credit_map": [ + 0.005477613769471645, + 0.007079923525452614, + 0.006389145739376545, + 0.007014949806034565, + 0.008856569416821003, + 0.009681383147835732, + 0.011916333809494972, + 0.007821309380233288, + 0.006153733469545841, + 0.006944118067622185, + 0.008541770279407501, + 0.006763198412954807, + 0.007317824754863977, + 0.007472679018974304, + 0.0059214564971625805, + 0.009665309451520443, + 0.013068826869130135, + 0.007761392742395401, + 0.020986538380384445, + 0.010618860833346844, + 0.008498621173202991, + 0.014170070178806782, + 0.00805368460714817, + 0.02354281023144722, + 0.011136229149997234, + 0.010652217082679272, + 0.013308636844158173, + 0.012817563489079475, + 0.00617976626381278, + 0.012884867377579212, + 0.008164345286786556, + 0.010293162427842617, + 0.006612158380448818, + 0.007480634842067957, + 0.00949372723698616, + 0.009099440649151802, + 0.00462642265483737, + 0.007828823290765285, + 0.011069012805819511, + 0.006339931394904852, + 0.0364316925406456, + 0.022025039419531822, + 0.07143820822238922, + 0.09482752531766891, + 0.011563451960682869, + 0.07013646513223648, + 0.040354736149311066, + 0.03740236908197403, + 0.022251566872000694, + 0.010205043479800224, + 0.027853259816765785, + 0.01709122210741043, + 0.0069572823122143745, + 0.02218141034245491, + 0.007534313015639782, + 0.02476631850004196, + 0.00309794582426548, + 0.0030417051166296005, + 0.010291713289916515, + 0.003941458649933338, + 0.0021079485304653645, + 0.0074016740545630455, + 0.0015759224770590663, + 0.0032355657313019037, + 0.00112386723048985, + 0.0021139869932085276, + 0.002831922611221671, + 0.0016718701226636767, + 0.0010893141152337193, + 0.0022022887133061886, + 0.0008455265779048204, + 0.002208741381764412, + 0.0030594803392887115, + 0.005665148142725229, + 0.013094211928546429, + 0.005185616668313742, + 0.005507071036845446, + 0.011361446231603622, + 0.0021104381885379553, + 0.004510112572461367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.08904840797185898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031929248943924904 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031929248943924904, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031929248943924904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06941873580217361, + 0.1008521020412445, + 0.23108220100402832, + 0.05212874710559845, + 0.028158310800790787, + 0.07773659378290176, + 0.06417104601860046, + 0.09728784114122391, + 0.06531763076782227, + 0.21384678781032562 + ], + "credit_map": [ + 0.009629543870687485, + 0.010267239063978195, + 0.012690051458775997, + 0.006181307602673769, + 0.005133111961185932, + 0.010639780201017857, + 0.00515309302136302, + 0.009724611416459084, + 0.008083553984761238, + 0.011531369760632515, + 0.009462827816605568, + 0.012223818339407444, + 0.013288401067256927, + 0.012960199266672134, + 0.020956946536898613, + 0.012344995513558388, + 0.027781441807746887, + 0.026066875085234642, + 0.04573114961385727, + 0.021356917917728424, + 0.020104048773646355, + 0.027594896033406258, + 0.01606258936226368, + 0.04638427868485451, + 0.006445929873734713, + 0.0044151535257697105, + 0.01056806743144989, + 0.005744118709117174, + 0.003766955342143774, + 0.006928738206624985, + 0.005451269913464785, + 0.00880851224064827, + 0.003954013809561729, + 0.0029036295600235462, + 0.004475102759897709, + 0.003959849011152983, + 0.0018640116322785616, + 0.003160277847200632, + 0.0032277314458042383, + 0.004613695666193962, + 0.008056328631937504, + 0.003984800074249506, + 0.017446747049689293, + 0.01073443703353405, + 0.0035472640302032232, + 0.0158561822026968, + 0.010341444052755833, + 0.007769390009343624, + 0.010024305433034897, + 0.005017580930143595, + 0.012751132249832153, + 0.007451238110661507, + 0.0033929927740246058, + 0.0089141009375453, + 0.004333458840847015, + 0.012286238372325897, + 0.0075982967391610146, + 0.008048785850405693, + 0.027820967137813568, + 0.011371903121471405, + 0.005548051558434963, + 0.023904185742139816, + 0.0055140662007033825, + 0.007481588516384363, + 0.004979757592082024, + 0.009882400743663311, + 0.014951065182685852, + 0.005441482178866863, + 0.003890288295224309, + 0.016563929617404938, + 0.0028010134119540453, + 0.006807692814618349, + 0.009599712677299976, + 0.011503864079713821, + 0.07349876314401627, + 0.007088749669492245, + 0.02454741857945919, + 0.07486560940742493, + 0.005114756990224123, + 0.007627918850630522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.0031929248943924904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780699387192726 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780699387192726, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780699387192726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06916644424200058, + 0.0665549710392952, + 0.08611853420734406, + 0.09022974222898483, + 0.06709523499011993, + 0.08622296899557114, + 0.024477001279592514, + 0.06874356418848038, + 0.03137098625302315, + 0.4100205898284912 + ], + "credit_map": [ + 0.008203594945371151, + 0.010279940441250801, + 0.009661437012255192, + 0.0080546410754323, + 0.008181421086192131, + 0.009578506462275982, + 0.0073629808612167835, + 0.007843918167054653, + 0.005595646798610687, + 0.006709917448461056, + 0.006556080188602209, + 0.008366165682673454, + 0.012391173280775547, + 0.008346468210220337, + 0.010175897739827633, + 0.008413621224462986, + 0.010863900184631348, + 0.008276841603219509, + 0.017635051161050797, + 0.007919631898403168, + 0.008939296007156372, + 0.01020688097923994, + 0.008014168590307236, + 0.01426275260746479, + 0.007460564374923706, + 0.005969009827822447, + 0.021120725199580193, + 0.009204726666212082, + 0.008170625194907188, + 0.009434530511498451, + 0.010450804606080055, + 0.018418751657009125, + 0.006852816324681044, + 0.009939553216099739, + 0.010478898882865906, + 0.011510557495057583, + 0.003825797000899911, + 0.009910237044095993, + 0.0071794418618083, + 0.007397923618555069, + 0.009817512705922127, + 0.004853845573961735, + 0.015326591208577156, + 0.015521716326475143, + 0.0036559095606207848, + 0.013760614208877087, + 0.011322230100631714, + 0.011964544653892517, + 0.003027254017069936, + 0.00207866495475173, + 0.0032067897263914347, + 0.0038796195294708014, + 0.0019531144062057137, + 0.0030350310262292624, + 0.003354100277647376, + 0.003942427691072226, + 0.005336127243936062, + 0.006759174168109894, + 0.016541490331292152, + 0.010578413493931293, + 0.0044652498327195644, + 0.011849303729832172, + 0.004791691899299622, + 0.00842210277915001, + 0.0020381861831992865, + 0.005890679080039263, + 0.006167389452457428, + 0.003589665051549673, + 0.0023338650353252888, + 0.006121077574789524, + 0.0023222409654408693, + 0.002907878952100873, + 0.0191794540733099, + 0.02713518775999546, + 0.12673036754131317, + 0.016271641477942467, + 0.06184426695108414, + 0.13048525154590607, + 0.01278596930205822, + 0.015588430687785149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.009780699387192726 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion, this extreme close-up, tilting downward, follows a gleaming metal screw as it spirals through the air. The screw, a polished stainless steel Phillips head, tumbles gracefully, its meticulously machined threads catching a sharp, directional light from an unseen source above. This harsh light creates a mesmerizing display of brilliant, specular highlights that propagate along its metallic form as it rotates steadily about its longitudinal axis. The background is a dark, severely out-of-focus industrial space rendered in muted charcoal and dark brown, with motes of dust drifting lazily, catching the light and becoming soft, glowing bokeh balls. The dominant color palette is a high-contrast blend of the screw's cool, brilliant silver against the warm, dark void. The entire scene is imbued with a focused, almost poetic atmosphere, transforming the simple act of falling into a captivating ballet of light and movement into a captivating melody of light and movement.", + "seed": 10881, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_3243_seed_10881_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029737742245197298, + "vlm_logprob": -0.029737742245197298, + "vlm_yes_no_margin": 3.799999952316284, + "vlm_token_credit": -0.029737742245197298, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06088656559586525 + ], + "opposite_logprob_tokens": [ + -2.9358866214752197 + ], + "logprob_reward": -0.06088656559586525, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06088656559586525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2493705153465271, + 0.11837486177682877, + 0.06278862059116364, + 0.08546245098114014, + 0.08002055436372757, + 0.0965571254491806, + 0.06969846785068512, + 0.07838425785303116, + 0.04477597773075104, + 0.11456717550754547 + ], + "credit_map": [ + 0.014254080131649971, + 0.01343110017478466, + 0.007823733612895012, + 0.0133827431127429, + 0.009048880077898502, + 0.019329626113176346, + 0.08560710400342941, + 0.08649323135614395, + 0.013634994626045227, + 0.016325820237398148, + 0.0168563574552536, + 0.014663326554000378, + 0.010896945372223854, + 0.010622571222484112, + 0.013048453256487846, + 0.022326380014419556, + 0.00793471746146679, + 0.0055342018604278564, + 0.010632633231580257, + 0.006022694520652294, + 0.00863531231880188, + 0.0042622569017112255, + 0.006799773778766394, + 0.012967029586434364, + 0.008968394249677658, + 0.012003578245639801, + 0.016799677163362503, + 0.015711845830082893, + 0.00690857321023941, + 0.010490692220628262, + 0.006369720678776503, + 0.00820996705442667, + 0.008922009728848934, + 0.012220882810652256, + 0.009655473753809929, + 0.01681198738515377, + 0.010315719060599804, + 0.007783943321555853, + 0.008362861350178719, + 0.005947672296315432, + 0.008456678129732609, + 0.01849696785211563, + 0.013256976380944252, + 0.013058681972324848, + 0.008045535534620285, + 0.015113329514861107, + 0.010673314332962036, + 0.009455633349716663, + 0.006398205645382404, + 0.02137172408401966, + 0.010523424483835697, + 0.008488123305141926, + 0.004011586308479309, + 0.005683423951268196, + 0.005265434738248587, + 0.007956544868648052, + 0.009189683012664318, + 0.013295052573084831, + 0.009928264655172825, + 0.010025713592767715, + 0.007884872145950794, + 0.010465318337082863, + 0.009773957543075085, + 0.007821387611329556, + 0.005626967176795006, + 0.006609526462852955, + 0.0079458924010396, + 0.004479793831706047, + 0.004148311447352171, + 0.004577574785798788, + 0.0026127274613827467, + 0.008775177411735058, + 0.011930500157177448, + 0.02100926637649536, + 0.015908200293779373, + 0.020620906725525856, + 0.011882330290973186, + 0.01294859778136015, + 0.009876192547380924, + 0.010391169227659702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.06088656559586525 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601926371455193 + ], + "opposite_logprob_tokens": [ + -3.895601987838745 + ], + "logprob_reward": -0.020601926371455193, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601926371455193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2586703598499298, + 0.09872009605169296, + 0.030476203188300133, + 0.17454354465007782, + 0.07375454902648926, + 0.15787549316883087, + 0.11321111768484116, + 0.0525345653295517, + 0.023653559386730194, + 0.016560470685362816 + ], + "credit_map": [ + 0.01434093527495861, + 0.014410443603992462, + 0.01386087667196989, + 0.012868981808423996, + 0.007542083039879799, + 0.019969960674643517, + 0.09422121196985245, + 0.08145587891340256, + 0.01328490674495697, + 0.011960115283727646, + 0.01455135177820921, + 0.006734340917319059, + 0.009282810613512993, + 0.009370546787977219, + 0.0124084847047925, + 0.02112753689289093, + 0.004400631412863731, + 0.0030733891762793064, + 0.005685195326805115, + 0.002835091669112444, + 0.004739508498460054, + 0.002907180693000555, + 0.0037642705719918013, + 0.0030709358397871256, + 0.015344925224781036, + 0.02771182358264923, + 0.033005282282829285, + 0.027954628691077232, + 0.016270896419882774, + 0.02196206897497177, + 0.014783317223191261, + 0.017510613426566124, + 0.008476519957184792, + 0.010449951514601707, + 0.008178037591278553, + 0.015469453297555447, + 0.010047944262623787, + 0.006856762338429689, + 0.007477714214473963, + 0.006798164919018745, + 0.014020910486578941, + 0.025824252516031265, + 0.021516678854823112, + 0.021109525114297867, + 0.015956005081534386, + 0.024737561121582985, + 0.019033212214708328, + 0.015677351504564285, + 0.009972980245947838, + 0.030029039829969406, + 0.01554165780544281, + 0.013991928659379482, + 0.007989302277565002, + 0.012758005410432816, + 0.009159681387245655, + 0.013768520206212997, + 0.005658478010445833, + 0.01202049758285284, + 0.007163641508668661, + 0.006275627296417952, + 0.004241967108100653, + 0.005941796116530895, + 0.005373847205191851, + 0.005858706776052713, + 0.004845363553613424, + 0.0033772492315620184, + 0.00477641774341464, + 0.0030332712922245264, + 0.00204644026234746, + 0.0015602050116285682, + 0.001293193199671805, + 0.0027214197907596827, + 0.0024603689089417458, + 0.0023709447123110294, + 0.002030173083767295, + 0.0026610607746988535, + 0.0024127678479999304, + 0.0016056208405643702, + 0.0013817353174090385, + 0.0016377998981624842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.020601926371455193 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306339845061302 + ], + "opposite_logprob_tokens": [ + -3.1680634021759033 + ], + "logprob_reward": -0.04306339845061302, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306339845061302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07142793387174606, + 0.0650205984711647, + 0.03547283262014389, + 0.17484314739704132, + 0.15021589398384094, + 0.3022027611732483, + 0.13564012944698334, + 0.02251317724585533, + 0.02721995674073696, + 0.015443607233464718 + ], + "credit_map": [ + 0.007310724817216396, + 0.006085114553570747, + 0.005953262094408274, + 0.005774200428277254, + 0.004468544851988554, + 0.006130135618150234, + 0.017610564827919006, + 0.018095385283231735, + 0.006373904645442963, + 0.009864719584584236, + 0.008745815604925156, + 0.007778450846672058, + 0.005866466090083122, + 0.008058147504925728, + 0.007510867901146412, + 0.010822214186191559, + 0.004637461621314287, + 0.003576868213713169, + 0.005803141742944717, + 0.0036908420734107494, + 0.004378591198474169, + 0.0031709345057606697, + 0.0032911712769418955, + 0.00692381476983428, + 0.021598538383841515, + 0.025572897866368294, + 0.02764500118792057, + 0.025560250505805016, + 0.015504456125199795, + 0.025317661464214325, + 0.01524373609572649, + 0.018400592729449272, + 0.01487824134528637, + 0.023595619946718216, + 0.016936229541897774, + 0.027378270402550697, + 0.026446763426065445, + 0.01332088466733694, + 0.014629272744059563, + 0.01303059235215187, + 0.023609550669789314, + 0.047854579985141754, + 0.03798384591937065, + 0.04082838445901871, + 0.03380570188164711, + 0.05158444121479988, + 0.03750568628311157, + 0.029030518606305122, + 0.012143229134380817, + 0.0351712629199028, + 0.019812095910310745, + 0.017194613814353943, + 0.0071899923495948315, + 0.01624302566051483, + 0.012166431173682213, + 0.01571946032345295, + 0.002547269919887185, + 0.0035806386731565, + 0.002855944447219372, + 0.0029998873360455036, + 0.002225418109446764, + 0.003278068732470274, + 0.002703848062083125, + 0.0023221003357321024, + 0.003427104791626334, + 0.004145625047385693, + 0.006769942119717598, + 0.003499759128317237, + 0.002043185057118535, + 0.001828930457122624, + 0.0015023540472611785, + 0.00400305213406682, + 0.0019527736585587263, + 0.0019839955493807793, + 0.0031098888721317053, + 0.0024516042321920395, + 0.001504150335676968, + 0.0015283389948308468, + 0.0012112028198316693, + 0.001701652305200696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.04306339845061302 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948696285486221 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948696285486221, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948696285486221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09783528000116348, + 0.0813513845205307, + 0.07473810762166977, + 0.23279505968093872, + 0.1573711484670639, + 0.11470381170511246, + 0.09869030863046646, + 0.05809199810028076, + 0.031500983983278275, + 0.05292188748717308 + ], + "credit_map": [ + 0.008864123374223709, + 0.014149235561490059, + 0.008851961232721806, + 0.008886463940143585, + 0.00611878139898181, + 0.009154774248600006, + 0.016314787790179253, + 0.025495152920484543, + 0.009626062586903572, + 0.006944356951862574, + 0.011681860312819481, + 0.0070104883052408695, + 0.007078224327415228, + 0.007554896175861359, + 0.01037567388266325, + 0.02107982151210308, + 0.013472235761582851, + 0.00422216160222888, + 0.01452825590968132, + 0.009948953986167908, + 0.006487378850579262, + 0.0058610145933926105, + 0.006736507639288902, + 0.013481593690812588, + 0.03368628770112991, + 0.02434118278324604, + 0.0338168703019619, + 0.03619527816772461, + 0.021933967247605324, + 0.03698589280247688, + 0.016085198149085045, + 0.02975037507712841, + 0.015796350315213203, + 0.024151809513568878, + 0.02117845229804516, + 0.029473356902599335, + 0.021552080288529396, + 0.017073743045330048, + 0.014388187788426876, + 0.013757176697254181, + 0.012601844035089016, + 0.02101459540426731, + 0.016597941517829895, + 0.019541988149285316, + 0.007830158807337284, + 0.014750928618013859, + 0.010554950684309006, + 0.01181140448898077, + 0.009145072661340237, + 0.020007453858852386, + 0.01786835864186287, + 0.011014064773917198, + 0.006504726596176624, + 0.014090661890804768, + 0.009109304286539555, + 0.01095067523419857, + 0.007430652622133493, + 0.012547779828310013, + 0.006967674009501934, + 0.0067679318599402905, + 0.0054516661912202835, + 0.006003811489790678, + 0.005676703527569771, + 0.007245777640491724, + 0.004849436692893505, + 0.0057159955613315105, + 0.006523550022393465, + 0.004610926378518343, + 0.0023166958708316088, + 0.0019015815341845155, + 0.001867096871137619, + 0.003715702798217535, + 0.007928437553346157, + 0.008244536817073822, + 0.009735497646033764, + 0.008217837661504745, + 0.006339170038700104, + 0.004444329533725977, + 0.0029254548717290163, + 0.005086622200906277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.005948696285486221 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188124522566795 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.018188124522566795, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018188124522566795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12275394797325134, + 0.0510236956179142, + 0.061470482498407364, + 0.2798188030719757, + 0.16463352739810944, + 0.09279289841651917, + 0.06324047595262527, + 0.07846028357744217, + 0.03434294834733009, + 0.051462896168231964 + ], + "credit_map": [ + 0.0075073447078466415, + 0.008759303949773312, + 0.008542339317500591, + 0.009461605921387672, + 0.005213382188230753, + 0.00905104074627161, + 0.04150788486003876, + 0.032711055129766464, + 0.005449181888252497, + 0.004989701323211193, + 0.008580329827964306, + 0.004806111566722393, + 0.0064651439897716045, + 0.005411889869719744, + 0.004310925491154194, + 0.011010415852069855, + 0.007592584937810898, + 0.006631559692323208, + 0.010582206770777702, + 0.006065875291824341, + 0.007596859708428383, + 0.004343126900494099, + 0.006172177381813526, + 0.012486100196838379, + 0.04032914713025093, + 0.03374933451414108, + 0.04039734974503517, + 0.04804833605885506, + 0.023156924173235893, + 0.042563892900943756, + 0.02086326666176319, + 0.030710577964782715, + 0.02169431559741497, + 0.015244892798364162, + 0.02763284556567669, + 0.047334469854831696, + 0.012236940674483776, + 0.017877239733934402, + 0.010241287760436535, + 0.012371553108096123, + 0.008293528109788895, + 0.014960422180593014, + 0.016534879803657532, + 0.01361029502004385, + 0.007844104431569576, + 0.010511333122849464, + 0.011602305807173252, + 0.009436034597456455, + 0.006000364199280739, + 0.013431299477815628, + 0.011373150162398815, + 0.00758256483823061, + 0.004794204141944647, + 0.006437940523028374, + 0.007215000689029694, + 0.006405956577509642, + 0.009404661133885384, + 0.01767253503203392, + 0.010231893509626389, + 0.009161497466266155, + 0.007554938551038504, + 0.00806986540555954, + 0.006689171772450209, + 0.009675728157162666, + 0.004103921353816986, + 0.004361140541732311, + 0.006353000178933144, + 0.005501855630427599, + 0.003879990428686142, + 0.0030988515354692936, + 0.0032242920715361834, + 0.003819897072389722, + 0.005978228524327278, + 0.010064772330224514, + 0.005311061628162861, + 0.008936945348978043, + 0.005423079710453749, + 0.006426029838621616, + 0.004198851063847542, + 0.0051239305175840855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.018188124522566795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion, this extreme close-up, tilting downward, follows a gleaming metal screw as it spirals through the air. The screw, a polished stainless steel Phillips head, tumbles gracefully, its meticulously machined threads catching a sharp, directional light from an unseen source above. This harsh light creates a mesmerizing display of brilliant, specular highlights that propagate along its metallic form as it rotates steadily about its longitudinal axis. The background is a dark, severely out-of-focus industrial space rendered in muted charcoal and dark brown, with motes of dust drifting lazily, catching the light and becoming soft, glowing bokeh balls. The dominant color palette is a high-contrast blend of the screw's cool, brilliant silver against the warm, dark void. The entire scene is imbued with a focused, almost poetic atmosphere, transforming the simple act of falling into a captivating ballet of light and movement into a captivating melody of light and movement.", + "seed": 10882, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_3243_seed_10882_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05391849614679813, + "vlm_logprob": -0.05391849614679813, + "vlm_yes_no_margin": 3.8749999523162844, + "vlm_token_credit": -0.05391849614679813, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0234614759683609 + ], + "opposite_logprob_tokens": [ + -4.02346134185791 + ], + "logprob_reward": -0.0234614759683609, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0234614759683609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08809369057416916, + 0.05225220322608948, + 0.050395503640174866, + 0.07972124218940735, + 0.03833343833684921, + 0.08096420019865036, + 0.34340551495552063, + 0.0958007276058197, + 0.05476609990000725, + 0.11626745015382767 + ], + "credit_map": [ + 0.008206012658774853, + 0.009528707712888718, + 0.00819934718310833, + 0.007647163234651089, + 0.011316615156829357, + 0.020581519231200218, + 0.00815439224243164, + 0.014459924772381783, + 0.005498639307916164, + 0.004496527370065451, + 0.005001532845199108, + 0.003910382743924856, + 0.0055129192769527435, + 0.010611268691718578, + 0.007084970828145742, + 0.010135957039892673, + 0.004382154438644648, + 0.00470519857481122, + 0.0034155917819589376, + 0.003729162970557809, + 0.007952972315251827, + 0.009908674284815788, + 0.0056970748119056225, + 0.010604668408632278, + 0.005719760432839394, + 0.012643391266465187, + 0.00533931702375412, + 0.010476414114236832, + 0.01297897007316351, + 0.010142668150365353, + 0.012029877863824368, + 0.01039083767682314, + 0.005790770519524813, + 0.002583392895758152, + 0.007345140911638737, + 0.002953561255708337, + 0.00545044569298625, + 0.006895708851516247, + 0.0029801735654473305, + 0.004334240220487118, + 0.005360129754990339, + 0.014728090725839138, + 0.009360432624816895, + 0.007777092978358269, + 0.01289521623402834, + 0.014202363789081573, + 0.01123677846044302, + 0.005404083523899317, + 0.014752737246453762, + 0.009542460553348064, + 0.09777599573135376, + 0.015985090285539627, + 0.03647942468523979, + 0.13261620700359344, + 0.012498770840466022, + 0.023754792287945747, + 0.007786066271364689, + 0.00938795693218708, + 0.01835094578564167, + 0.007553591392934322, + 0.009182409383356571, + 0.02386801317334175, + 0.010107625275850296, + 0.009564108215272427, + 0.005359628237783909, + 0.010177765972912312, + 0.007449707482010126, + 0.006936084944754839, + 0.005891293752938509, + 0.008017849177122116, + 0.006475582718849182, + 0.004458182491362095, + 0.016298191621899605, + 0.017541293054819107, + 0.01831027865409851, + 0.01759987883269787, + 0.016323082149028778, + 0.006253769621253014, + 0.014010661281645298, + 0.009930282831192017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.0234614759683609 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372495621442795 + ], + "opposite_logprob_tokens": [ + -3.4087250232696533 + ], + "logprob_reward": -0.03372495621442795, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372495621442795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026267603039741516, + 0.039125628769397736, + 0.18391576409339905, + 0.04786230996251106, + 0.02342967689037323, + 0.06196656823158264, + 0.5190998911857605, + 0.04259813576936722, + 0.022588444873690605, + 0.033145930618047714 + ], + "credit_map": [ + 0.0029537349473685026, + 0.0024423019494861364, + 0.0025447586085647345, + 0.0025062731001526117, + 0.002255606697872281, + 0.005703295581042767, + 0.0020642178133130074, + 0.005797417834401131, + 0.003998544532805681, + 0.004674557596445084, + 0.0048549058847129345, + 0.0032060816884040833, + 0.0036802245303988457, + 0.007812727242708206, + 0.004271708894520998, + 0.00662688072770834, + 0.014578145928680897, + 0.0072450111620128155, + 0.0060151368379592896, + 0.010001220740377903, + 0.011216950602829456, + 0.057179972529411316, + 0.009380162693560123, + 0.06829915940761566, + 0.0047426787205040455, + 0.004064902663230896, + 0.0067709279246628284, + 0.005298336036503315, + 0.005331431981176138, + 0.006110594607889652, + 0.006652309559285641, + 0.00889112800359726, + 0.0029884413816034794, + 0.0025311799254268408, + 0.003249490400776267, + 0.002234117593616247, + 0.003608020721003413, + 0.003881619544699788, + 0.002808730350807309, + 0.002128077205270529, + 0.006507238373160362, + 0.008889704011380672, + 0.010257486253976822, + 0.005185646470636129, + 0.007845818996429443, + 0.009786758571863174, + 0.007545726839452982, + 0.005948186852037907, + 0.023153718560934067, + 0.015879251062870026, + 0.14424213767051697, + 0.02960028313100338, + 0.05333923175930977, + 0.19669431447982788, + 0.02250087447464466, + 0.03369011729955673, + 0.0030078929848968983, + 0.00564472284168005, + 0.00864572823047638, + 0.0027044550515711308, + 0.0042845276184380054, + 0.011502392590045929, + 0.0031520635820925236, + 0.0036563556641340256, + 0.003211854724213481, + 0.0028835032135248184, + 0.0029716319404542446, + 0.003217860125005245, + 0.0022351243533194065, + 0.002897914033383131, + 0.0032754107378423214, + 0.0018951456295326352, + 0.005227672401815653, + 0.005385055206716061, + 0.004976881667971611, + 0.004067712929099798, + 0.005537336226552725, + 0.001766760484315455, + 0.0030825205612927675, + 0.0031019924208521843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.03372495621442795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157258212566376 + ], + "opposite_logprob_tokens": [ + -1.7015725374221802 + ], + "logprob_reward": -0.20157258212566376, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157258212566376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05024528130888939, + 0.035406313836574554, + 0.05922940373420715, + 0.016565373167395592, + 0.01999664306640625, + 0.05228482559323311, + 0.5898269414901733, + 0.11445789039134979, + 0.0415363535284996, + 0.020450973883271217 + ], + "credit_map": [ + 0.0038698920980095863, + 0.004362713545560837, + 0.003424151800572872, + 0.0028190654702484608, + 0.006751338019967079, + 0.013027709908783436, + 0.0034525571390986443, + 0.01253785751760006, + 0.003293232759460807, + 0.003245169296860695, + 0.0028733545914292336, + 0.0028914217837154865, + 0.004605909809470177, + 0.006312247831374407, + 0.0048007722944021225, + 0.007384208496659994, + 0.00511091947555542, + 0.0035787581000477076, + 0.003481960855424404, + 0.0035069298464804888, + 0.005659549497067928, + 0.014045337215065956, + 0.0052932132966816425, + 0.018552739173173904, + 0.0017595614772289991, + 0.001676518702879548, + 0.0018372505437582731, + 0.0017640271689742804, + 0.0020748800598084927, + 0.0025079615879803896, + 0.0018193209543824196, + 0.0031258538365364075, + 0.001983978785574436, + 0.0024513360112905502, + 0.0025562793016433716, + 0.0019351659575477242, + 0.0030349998269230127, + 0.0029053043108433485, + 0.0029815754387527704, + 0.002148003550246358, + 0.0053067198023200035, + 0.005657713860273361, + 0.0060642617754638195, + 0.005051703192293644, + 0.009530631825327873, + 0.010980503633618355, + 0.0061486586928367615, + 0.00354463467374444, + 0.03629421815276146, + 0.02336275763809681, + 0.15122424066066742, + 0.02776193805038929, + 0.07008537650108337, + 0.20582453906536102, + 0.02671002596616745, + 0.04856386408209801, + 0.00904788263142109, + 0.01533486694097519, + 0.024207711219787598, + 0.005747124552726746, + 0.0109377671033144, + 0.03257448226213455, + 0.005498620215803385, + 0.011109449900686741, + 0.005343522876501083, + 0.009172762744128704, + 0.0054181041195988655, + 0.004562816582620144, + 0.00345939421094954, + 0.004557080566883087, + 0.005134884733706713, + 0.0038877904880791903, + 0.002760315081104636, + 0.003049638122320175, + 0.002928453730419278, + 0.0028083850629627705, + 0.002915126970037818, + 0.0015273385215550661, + 0.0024742092937231064, + 0.001987508963793516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.20157258212566376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095619544386864 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095619544386864, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095619544386864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0762176439166069, + 0.10454968363046646, + 0.2671947181224823, + 0.07230193167924881, + 0.12161063402891159, + 0.10550056397914886, + 0.0255403034389019, + 0.05805101618170738, + 0.02517152763903141, + 0.14386199414730072 + ], + "credit_map": [ + 0.0066864825785160065, + 0.006478381343185902, + 0.00914224237203598, + 0.006335179787129164, + 0.007698429748415947, + 0.012884452939033508, + 0.006617771927267313, + 0.020374702289700508, + 0.010346870869398117, + 0.006814612075686455, + 0.00922494288533926, + 0.007555944845080376, + 0.006186123937368393, + 0.02616872638463974, + 0.011816535145044327, + 0.026435932144522667, + 0.02231602370738983, + 0.01019161194562912, + 0.019344663247466087, + 0.012833407148718834, + 0.013959026895463467, + 0.07817867398262024, + 0.018422290682792664, + 0.09194903075695038, + 0.005657763220369816, + 0.007833708077669144, + 0.0072165788151323795, + 0.010924349538981915, + 0.011524129658937454, + 0.009382177144289017, + 0.011720494367182255, + 0.008042723871767521, + 0.015473369508981705, + 0.010490627028048038, + 0.021558772772550583, + 0.009592676535248756, + 0.018788371235132217, + 0.020144399255514145, + 0.011558824218809605, + 0.014003594405949116, + 0.009944229386746883, + 0.014944114722311497, + 0.016737399622797966, + 0.007651061750948429, + 0.016720512881875038, + 0.018928632140159607, + 0.013301907107234001, + 0.007272713351994753, + 0.0036543114110827446, + 0.0020019488874822855, + 0.0036276851315051317, + 0.0035303370095789433, + 0.00292104110121727, + 0.003483458422124386, + 0.0032821809872984886, + 0.0030393428169190884, + 0.00470538716763258, + 0.003810061141848564, + 0.01260487362742424, + 0.00388203258626163, + 0.004875193815678358, + 0.01779518648982048, + 0.0035980050452053547, + 0.006780276075005531, + 0.003524631028994918, + 0.003003880847245455, + 0.003566830651834607, + 0.0029542504344135523, + 0.001726715941913426, + 0.0036838010419160128, + 0.003614056156948209, + 0.0030973630491644144, + 0.021398911252617836, + 0.014597434550523758, + 0.029523761942982674, + 0.00984319020062685, + 0.013867992907762527, + 0.01243588887155056, + 0.02297361195087433, + 0.019221201539039612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.004095619544386864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737846881151199 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737846881151199, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737846881151199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08381972461938858, + 0.11874964833259583, + 0.07567762583494186, + 0.04012168571352959, + 0.10795911401510239, + 0.09162332117557526, + 0.3099804222583771, + 0.05846942216157913, + 0.050108976662158966, + 0.06349007785320282 + ], + "credit_map": [ + 0.0094510642811656, + 0.007191317621618509, + 0.01420113630592823, + 0.008853878825902939, + 0.00794835202395916, + 0.013606410473585129, + 0.00694633973762393, + 0.01562122069299221, + 0.010644433088600636, + 0.007962248288094997, + 0.010218890383839607, + 0.0057932185009121895, + 0.007610229775309563, + 0.03606422245502472, + 0.008704615756869316, + 0.03175177052617073, + 0.007943079806864262, + 0.0037420447915792465, + 0.010401665233075619, + 0.005596899427473545, + 0.006669516675174236, + 0.016367599368095398, + 0.007214711047708988, + 0.017742108553647995, + 0.0042974757961928844, + 0.0031780966091901064, + 0.00764345470815897, + 0.004023420624434948, + 0.004394788760691881, + 0.005491175688803196, + 0.005318475421518087, + 0.005774795077741146, + 0.01590764708817005, + 0.009298481978476048, + 0.02318652532994747, + 0.009184004738926888, + 0.014101437292993069, + 0.01645696349442005, + 0.008238951675593853, + 0.011585096828639507, + 0.0068341768346726894, + 0.008742701262235641, + 0.01255402248352766, + 0.010942934080958366, + 0.01210224349051714, + 0.016524137929081917, + 0.014313844032585621, + 0.00960924569517374, + 0.014204010367393494, + 0.013744929805397987, + 0.07982494682073593, + 0.03105401247739792, + 0.02974449098110199, + 0.10722706466913223, + 0.019407162442803383, + 0.014773795381188393, + 0.004574028309434652, + 0.0052803270518779755, + 0.010678302496671677, + 0.007039390038698912, + 0.005715751089155674, + 0.013683225028216839, + 0.0056754956021904945, + 0.005822900217026472, + 0.006593675818294287, + 0.0075728027150034904, + 0.006074188277125359, + 0.00993918627500534, + 0.00505400076508522, + 0.004372365307062864, + 0.0062381913885474205, + 0.004264560528099537, + 0.008867510594427586, + 0.010845966637134552, + 0.009946929290890694, + 0.009259499609470367, + 0.007953899912536144, + 0.00402619456872344, + 0.0075983754359185696, + 0.004991692025214434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.006737846881151199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in dramatic slow motion, this extreme close-up, tilting downward, follows a gleaming metal screw as it spirals through the air. The screw, a polished stainless steel Phillips head, tumbles gracefully, its meticulously machined threads catching a sharp, directional light from an unseen source above. This harsh light creates a mesmerizing display of brilliant, specular highlights that propagate along its metallic form as it rotates steadily about its longitudinal axis. The background is a dark, severely out-of-focus industrial space rendered in muted charcoal and dark brown, with motes of dust drifting lazily, catching the light and becoming soft, glowing bokeh balls. The dominant color palette is a high-contrast blend of the screw's cool, brilliant silver against the warm, dark void. The entire scene is imbued with a focused, almost poetic atmosphere, transforming the simple act of falling into a captivating ballet of light and movement into a captivating melody of light and movement.", + "seed": 10883, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_7/index_3243_seed_10883_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08356347568333149, + "vlm_logprob": -0.08356347568333149, + "vlm_yes_no_margin": 3.8999999523162843, + "vlm_token_credit": -0.08356347568333149, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024144308641552925 + ], + "opposite_logprob_tokens": [ + -4.024144172668457 + ], + "logprob_reward": -0.024144308641552925, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.024144308641552925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17903921008110046, + 0.0992346778512001, + 0.055252283811569214, + 0.0792238637804985, + 0.07593294978141785, + 0.051798220723867416, + 0.11146584898233414, + 0.2570257782936096, + 0.0592016726732254, + 0.031825482845306396 + ], + "credit_map": [ + 0.01833445206284523, + 0.028421029448509216, + 0.022552229464054108, + 0.0228200014680624, + 0.023630671203136444, + 0.028770102187991142, + 0.019951388239860535, + 0.01455932017415762, + 0.009769001044332981, + 0.008141770027577877, + 0.02222244255244732, + 0.0076169585809111595, + 0.006854038219898939, + 0.016655225306749344, + 0.006811571307480335, + 0.021163661032915115, + 0.008380935527384281, + 0.0035949007142335176, + 0.012762059457600117, + 0.004492342006415129, + 0.003907598555088043, + 0.010401960462331772, + 0.004883190616965294, + 0.006829299032688141, + 0.007321062032133341, + 0.007898065261542797, + 0.01925789937376976, + 0.008392004296183586, + 0.005026164930313826, + 0.015709910541772842, + 0.008907773531973362, + 0.006710980087518692, + 0.00872450228780508, + 0.007895879447460175, + 0.010940569452941418, + 0.007002514321357012, + 0.007697187829762697, + 0.00977352075278759, + 0.010336486622691154, + 0.013562285341322422, + 0.005038802977651358, + 0.00531249912455678, + 0.007287760730832815, + 0.004579466767609119, + 0.0047545647248625755, + 0.007391735911369324, + 0.007735567167401314, + 0.00969782005995512, + 0.012619621120393276, + 0.010048174299299717, + 0.024142447859048843, + 0.007895054295659065, + 0.006174721755087376, + 0.009792054072022438, + 0.014795327559113503, + 0.025998450815677643, + 0.04159068316221237, + 0.02606962062418461, + 0.028317809104919434, + 0.02132711373269558, + 0.03946061432361603, + 0.03891254588961601, + 0.03202980011701584, + 0.029317589476704597, + 0.008256829343736172, + 0.005271370522677898, + 0.007063237018883228, + 0.004362098872661591, + 0.008405798114836216, + 0.011042817495763302, + 0.004478642251342535, + 0.010320878587663174, + 0.003238484263420105, + 0.003001568838953972, + 0.0032601396087557077, + 0.0027158530429005623, + 0.005978900007903576, + 0.0045913150534033775, + 0.004641255829483271, + 0.00439796457067132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.024144308641552925 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381210558116436 + ], + "opposite_logprob_tokens": [ + -3.288120985031128 + ], + "logprob_reward": -0.0381210558116436, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381210558116436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12490050494670868, + 0.05950025096535683, + 0.04473216459155083, + 0.09233149141073227, + 0.08489573001861572, + 0.054951950907707214, + 0.1619827002286911, + 0.1321706920862198, + 0.14497748017311096, + 0.09955697506666183 + ], + "credit_map": [ + 0.013894974254071712, + 0.018880389630794525, + 0.01704374887049198, + 0.01605258509516716, + 0.016691602766513824, + 0.019125493243336678, + 0.012285307981073856, + 0.010926416143774986, + 0.007328320760279894, + 0.005802788771688938, + 0.01081869937479496, + 0.006937747355550528, + 0.005287554115056992, + 0.009208789095282555, + 0.006429922301322222, + 0.007686436176300049, + 0.006809577811509371, + 0.0039825658313930035, + 0.007997131906449795, + 0.005636161658912897, + 0.003697645151987672, + 0.0071309772320091724, + 0.0043199313804507256, + 0.005158174782991409, + 0.01042111311107874, + 0.005295755807310343, + 0.017648175358772278, + 0.004714461974799633, + 0.004236598499119282, + 0.01843906193971634, + 0.006181120406836271, + 0.02539519965648651, + 0.012069781310856342, + 0.008727495558559895, + 0.011505196802318096, + 0.011132915504276752, + 0.007798874285072088, + 0.010814174078404903, + 0.009537572972476482, + 0.013309724628925323, + 0.006167492363601923, + 0.004908316768705845, + 0.00932990200817585, + 0.004389591049402952, + 0.00454992800951004, + 0.008814043365418911, + 0.007152331992983818, + 0.009640351869165897, + 0.021168483421206474, + 0.026346374303102493, + 0.015690747648477554, + 0.010544590651988983, + 0.019683634862303734, + 0.024837976321578026, + 0.02395337074995041, + 0.019757535308599472, + 0.02167397364974022, + 0.01441920269280672, + 0.014860178343951702, + 0.009762871079146862, + 0.02251427061855793, + 0.017171937972307205, + 0.018245702609419823, + 0.013522556982934475, + 0.014644560404121876, + 0.01978171244263649, + 0.018319984897971153, + 0.014316602610051632, + 0.022702233865857124, + 0.019745459780097008, + 0.018526803702116013, + 0.016940124332904816, + 0.010296142660081387, + 0.0150142852216959, + 0.010441161692142487, + 0.008523024618625641, + 0.013208399526774883, + 0.007976982742547989, + 0.014799633994698524, + 0.01929735392332077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.0381210558116436 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486380875110626 + ], + "opposite_logprob_tokens": [ + -1.2236380577087402 + ], + "logprob_reward": -0.3486380875110626, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486380875110626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22529758512973785, + 0.09924688935279846, + 0.07294313609600067, + 0.05286993458867073, + 0.026089858263731003, + 0.030475212261080742, + 0.11724783480167389, + 0.20578685402870178, + 0.12381909787654877, + 0.04622356966137886 + ], + "credit_map": [ + 0.028049923479557037, + 0.03180554509162903, + 0.0309918113052845, + 0.029969947412610054, + 0.026416189968585968, + 0.03334422409534454, + 0.021337032318115234, + 0.023382918909192085, + 0.008585306815803051, + 0.007655580993741751, + 0.01758827641606331, + 0.009790632873773575, + 0.009624259546399117, + 0.014458205550909042, + 0.010902294889092445, + 0.02064233273267746, + 0.013617980293929577, + 0.006627623923122883, + 0.014589378610253334, + 0.006851824466139078, + 0.006029821000993252, + 0.011069424450397491, + 0.005374558735638857, + 0.008782530203461647, + 0.006021131295710802, + 0.005022067576646805, + 0.009454105980694294, + 0.0025146016851067543, + 0.0030170432291924953, + 0.012367633171379566, + 0.003242089645937085, + 0.011231260374188423, + 0.003634322201833129, + 0.0029243428725749254, + 0.004207818303257227, + 0.00336865964345634, + 0.0022123747039586306, + 0.00273845624178648, + 0.0027974825352430344, + 0.004206402227282524, + 0.003475793404504657, + 0.00427437387406826, + 0.004540367051959038, + 0.0025141453370451927, + 0.0033269161358475685, + 0.004088672809302807, + 0.003425290109589696, + 0.004829653073102236, + 0.0136489886790514, + 0.012365581467747688, + 0.026181383058428764, + 0.011399921961128712, + 0.010793430730700493, + 0.010247421450912952, + 0.01000541914254427, + 0.022605683654546738, + 0.025573190301656723, + 0.021448230370879173, + 0.019732987508177757, + 0.01742524281144142, + 0.038225919008255005, + 0.024843113496899605, + 0.023447109386324883, + 0.03509105369448662, + 0.012518912553787231, + 0.012679697945713997, + 0.01595192775130272, + 0.00909015629440546, + 0.015454461798071861, + 0.025884512811899185, + 0.013527448289096355, + 0.01871197670698166, + 0.0039333500899374485, + 0.008546525612473488, + 0.005298321135342121, + 0.0035795350559055805, + 0.007914184592664242, + 0.0051646302454173565, + 0.004859158769249916, + 0.006927866954356432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.3486380875110626 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094195086508989 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094195086508989, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094195086508989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14814642071723938, + 0.12768371403217316, + 0.06215251237154007, + 0.11357180774211884, + 0.07494993507862091, + 0.02887549065053463, + 0.09001519531011581, + 0.09689144790172577, + 0.161594957113266, + 0.0961185097694397 + ], + "credit_map": [ + 0.01646198332309723, + 0.017789194360375404, + 0.025361323729157448, + 0.020411387085914612, + 0.014958971180021763, + 0.022650590166449547, + 0.01515213679522276, + 0.015360834077000618, + 0.016588596627116203, + 0.00859849713742733, + 0.0249998327344656, + 0.010073157027363777, + 0.007651264779269695, + 0.019843822345137596, + 0.010769653134047985, + 0.02915889583528042, + 0.010666961781680584, + 0.0042814877815544605, + 0.012651474215090275, + 0.005074439104646444, + 0.004561966750770807, + 0.009985935874283314, + 0.006298330146819353, + 0.00863192044198513, + 0.018671242520213127, + 0.01024110522121191, + 0.018457384780049324, + 0.007888918742537498, + 0.004361256025731564, + 0.014340403489768505, + 0.01159308385103941, + 0.02801840379834175, + 0.010423130355775356, + 0.006957868579775095, + 0.010497014038264751, + 0.009358542039990425, + 0.006793520413339138, + 0.008962173946201801, + 0.007893617264926434, + 0.014064076356589794, + 0.002885989146307111, + 0.002822907641530037, + 0.004062579479068518, + 0.002880791202187538, + 0.0031755126547068357, + 0.00360146164894104, + 0.004645719192922115, + 0.0048005301505327225, + 0.012198377400636673, + 0.010011815465986729, + 0.010882008820772171, + 0.008489841595292091, + 0.009282522834837437, + 0.012443304061889648, + 0.013249099254608154, + 0.01345822773873806, + 0.013167041353881359, + 0.011143538169562817, + 0.009585747495293617, + 0.008796517737209797, + 0.01729423552751541, + 0.01240403950214386, + 0.01075559388846159, + 0.01374473050236702, + 0.018610911443829536, + 0.015699679031968117, + 0.020851250737905502, + 0.012075725942850113, + 0.02445623092353344, + 0.031032707542181015, + 0.012712949886918068, + 0.0261555016040802, + 0.010027502663433552, + 0.008107529021799564, + 0.012919743545353413, + 0.006225121207535267, + 0.010310174897313118, + 0.024737291038036346, + 0.0098410714417696, + 0.01395008247345686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.004094195086508989 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gleaming metal screw with a polished stainless steel Phillips head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the screw shown tumbling and rotating steadily about its longitudinal axis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene shot as an extreme close-up with a camera tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, severely out-of-focus industrial space with drifting dust motes appearing as soft glowing bokeh balls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature sharp, directional lighting creating brilliant specular highlights on the screw's threads?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819731365889311 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819731365889311, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819731365889311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13969546556472778, + 0.14705491065979004, + 0.0774756669998169, + 0.02176435850560665, + 0.06762966513633728, + 0.05898439884185791, + 0.15128076076507568, + 0.2195536196231842, + 0.053385522216558456, + 0.06317559629678726 + ], + "credit_map": [ + 0.014846683479845524, + 0.014790653251111507, + 0.024224644526839256, + 0.02065179869532585, + 0.013660203665494919, + 0.022641252726316452, + 0.01345317717641592, + 0.0154270576313138, + 0.020394835621118546, + 0.008159873075783253, + 0.03698578476905823, + 0.011236149817705154, + 0.008932632394134998, + 0.024318711832165718, + 0.008960199542343616, + 0.02806672267615795, + 0.01493768859654665, + 0.0041265287436544895, + 0.020125649869441986, + 0.0056248935870826244, + 0.004834933206439018, + 0.013384762220084667, + 0.004714107606559992, + 0.009727110154926777, + 0.0034378247801214457, + 0.0016505691455677152, + 0.003582201199606061, + 0.0025584641844034195, + 0.0018036534311249852, + 0.002577515784651041, + 0.0026992319617420435, + 0.003454898251220584, + 0.006794434040784836, + 0.008741512894630432, + 0.009270363487303257, + 0.007633875124156475, + 0.007014177739620209, + 0.007283929269760847, + 0.009043637663125992, + 0.011847742833197117, + 0.005706570111215115, + 0.005243763793259859, + 0.011481319554150105, + 0.0042059579864144325, + 0.004536909516900778, + 0.0074147689156234264, + 0.008255275897681713, + 0.012139835394918919, + 0.018109625205397606, + 0.015542905777692795, + 0.023739522323012352, + 0.01390120666474104, + 0.010132066905498505, + 0.01425252016633749, + 0.025562334805727005, + 0.030040588229894638, + 0.038349103182554245, + 0.020571550354361534, + 0.022652018815279007, + 0.02027728222310543, + 0.03153899684548378, + 0.03215937316417694, + 0.027291378006339073, + 0.02671390399336815, + 0.005768829490989447, + 0.005656233988702297, + 0.0063476162031292915, + 0.0050797369331121445, + 0.0058008949272334576, + 0.010764503851532936, + 0.005978669039905071, + 0.007989036850631237, + 0.004742471035569906, + 0.008963299915194511, + 0.005514295771718025, + 0.004443299025297165, + 0.015577102079987526, + 0.009124728851020336, + 0.0060086301527917385, + 0.008801770396530628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.002819731365889311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_8/index_tensor([2716, 2716])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_8/index_tensor([2716, 2716])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6ea34edc982263c6e457172105722f572a236481 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_8/index_tensor([2716, 2716])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning high-angle wide shot, as if shot from a drone's perspective, begins with a slow, steady forward motion, gliding across a tranquil, mirror-like lake that fills the foreground. In the distance, a majestic volcano dominates the frame, its massive mass rising from a lush landscape blanketed by dense, dark green forest. The volcano's summit is partially shrouded in thick, swirling white clouds, while a powerful plume of gray steam and black smoke continuously billows from its crater, creating a stunning contrast against the vast, deep blue sky. As the camera continues its deliberate advance, grand scale gives way to intricate detail; the rugged texture of the dark volcanic rock becomes visible, revealing a complex tapestry of jagged edges and ancient fissures. Occasionally, streams of incandescent lava slowly and deliberately flow down the sides, emitting a fiery orange-red glow that penetrates the dark rock and emits a subtle heat haze, emphasizing the scene's powerful aura of raw, untamed power and sublime natural beauty.", + "seed": 12640, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_8/index_2716_seed_12640_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06436337577179074, + "vlm_logprob": -0.06436337577179074, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.06436337577179074, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04959576204419136 + ], + "opposite_logprob_tokens": [ + -4.174595832824707 + ], + "logprob_reward": -0.04959576204419136, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.04959576204419136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1345091164112091, + 0.09592659771442413, + 0.059942469000816345, + 0.08059391379356384, + 0.036347515881061554, + 0.07397880405187607, + 0.1252513974905014, + 0.28413650393486023, + 0.08409905433654785, + 0.025214577093720436 + ], + "credit_map": [ + 0.012773039750754833, + 0.011614663526415825, + 0.02004244178533554, + 0.01730225421488285, + 0.01542262639850378, + 0.012893528677523136, + 0.018384773284196854, + 0.026075787842273712, + 0.011690584942698479, + 0.005443368572741747, + 0.005477281752973795, + 0.007780984044075012, + 0.016880938783288002, + 0.019796596840023994, + 0.014149307273328304, + 0.014707539230585098, + 0.008103166706860065, + 0.0048448797315359116, + 0.00439095264300704, + 0.007905879057943821, + 0.005501278210431337, + 0.007138094399124384, + 0.007919949479401112, + 0.014138272032141685, + 0.008959312923252583, + 0.00764745008200407, + 0.005802642088383436, + 0.0089295394718647, + 0.014776777476072311, + 0.009242726489901543, + 0.014698234386742115, + 0.010537233203649521, + 0.003430024953559041, + 0.004404144361615181, + 0.002102256752550602, + 0.0047415620647370815, + 0.0062054796144366264, + 0.004192625172436237, + 0.007117369212210178, + 0.004154051188379526, + 0.006211739964783192, + 0.00380814541131258, + 0.006006257142871618, + 0.013660997152328491, + 0.009679272770881653, + 0.007088633254170418, + 0.0144848907366395, + 0.013038869015872478, + 0.027487797662615776, + 0.007144746836274862, + 0.005444053560495377, + 0.011593922041356564, + 0.02241544984281063, + 0.015943793579936028, + 0.008825210854411125, + 0.026396416127681732, + 0.024520497769117355, + 0.027470815926790237, + 0.022059906274080276, + 0.0325281023979187, + 0.0562472864985466, + 0.05091550573706627, + 0.032168373465538025, + 0.03822603076696396, + 0.009550891816616058, + 0.010084046050906181, + 0.005477267317473888, + 0.007495669648051262, + 0.021102553233504295, + 0.015454811044037342, + 0.0064171855337917805, + 0.008516632951796055, + 0.00287526473402977, + 0.0024476719554513693, + 0.0016736815450713038, + 0.002697484800592065, + 0.0042540268041193485, + 0.0040787598118186, + 0.003106714691966772, + 0.004080973099917173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.04959576204419136 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062102172523736954 + ], + "opposite_logprob_tokens": [ + -2.8121020793914795 + ], + "logprob_reward": -0.062102172523736954, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062102172523736954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16187310218811035, + 0.18157240748405457, + 0.18399381637573242, + 0.09342879801988602, + 0.10961709171533585, + 0.08565815538167953, + 0.053268518298864365, + 0.06303679943084717, + 0.04424876719713211, + 0.023302560672163963 + ], + "credit_map": [ + 0.018957355991005898, + 0.019415944814682007, + 0.012912635691463947, + 0.01506821159273386, + 0.031123438850045204, + 0.027146972715854645, + 0.017739880830049515, + 0.01950865052640438, + 0.021202247589826584, + 0.015722373500466347, + 0.010520510375499725, + 0.015575939789414406, + 0.029966656118631363, + 0.032508909702301025, + 0.02204984799027443, + 0.03402591496706009, + 0.02845013700425625, + 0.00826428085565567, + 0.013672356493771076, + 0.02657703496515751, + 0.016832536086440086, + 0.014265158213675022, + 0.024604972451925278, + 0.051327336579561234, + 0.009403952397406101, + 0.00966392271220684, + 0.005230165086686611, + 0.009370701387524605, + 0.018826976418495178, + 0.01161240879446268, + 0.018169241026043892, + 0.01115142460912466, + 0.015453463420271873, + 0.009225049987435341, + 0.0075558326207101345, + 0.01497525442391634, + 0.008351610973477364, + 0.00606897845864296, + 0.021516377106308937, + 0.026470523327589035, + 0.007915610447525978, + 0.008250393904745579, + 0.007814515382051468, + 0.01737162470817566, + 0.0058769118040800095, + 0.005338392686098814, + 0.017974011600017548, + 0.015116688795387745, + 0.00866488367319107, + 0.005270715802907944, + 0.0036065946333110332, + 0.005488071125000715, + 0.008542396128177643, + 0.006300917826592922, + 0.003794141113758087, + 0.011600798927247524, + 0.006007387302815914, + 0.006336168386042118, + 0.005060484632849693, + 0.0066263931803405285, + 0.012115711346268654, + 0.010396798141300678, + 0.00713422242552042, + 0.009359633550047874, + 0.004742704331874847, + 0.005365996155887842, + 0.003731607925146818, + 0.007884395308792591, + 0.005110192112624645, + 0.0037096631713211536, + 0.006716547999531031, + 0.006987659726291895, + 0.002595061669126153, + 0.001557461335323751, + 0.0015667285770177841, + 0.0031589989084750414, + 0.0019118170021101832, + 0.002233478706330061, + 0.0035798600874841213, + 0.006699153687804937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.062102172523736954 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256047938019037 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256047938019037, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256047938019037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0552346408367157, + 0.18865366280078888, + 0.3194947838783264, + 0.09961027652025223, + 0.06384500116109848, + 0.12469112128019333, + 0.07348459959030151, + 0.037603747099637985, + 0.016408804804086685, + 0.02097330428659916 + ], + "credit_map": [ + 0.006453950423747301, + 0.009832068346440792, + 0.0046857800334692, + 0.008008476346731186, + 0.006891691125929356, + 0.0043054609559476376, + 0.006876809522509575, + 0.008180405013263226, + 0.020906684920191765, + 0.019673585891723633, + 0.008867058902978897, + 0.015687454491853714, + 0.03370947390794754, + 0.038026560097932816, + 0.024556780233979225, + 0.027226068079471588, + 0.05194413661956787, + 0.022962214425206184, + 0.020613612607121468, + 0.04726820066571236, + 0.028793448582291603, + 0.02341705933213234, + 0.033518798649311066, + 0.09097730368375778, + 0.007505256216973066, + 0.005923781543970108, + 0.006161675322800875, + 0.013821437023580074, + 0.008496997877955437, + 0.00843561626970768, + 0.03302554041147232, + 0.01623997464776039, + 0.00868671853095293, + 0.007322479970753193, + 0.003975555766373873, + 0.008740101009607315, + 0.005640567280352116, + 0.004571090918034315, + 0.010232292115688324, + 0.014676193706691265, + 0.00951471272855997, + 0.010367547161877155, + 0.011340293101966381, + 0.025571076199412346, + 0.007118181325495243, + 0.00602367427200079, + 0.031024204567074776, + 0.023731430992484093, + 0.01740352436900139, + 0.004063666798174381, + 0.004057878628373146, + 0.008478552103042603, + 0.008855700492858887, + 0.006293300073593855, + 0.006325262598693371, + 0.018006717786192894, + 0.0035087515134364367, + 0.004608809947967529, + 0.0030520607251673937, + 0.0047358060255646706, + 0.006527455057948828, + 0.004202934913337231, + 0.00496433163061738, + 0.00600359495729208, + 0.0017060062382370234, + 0.0016882454510778189, + 0.0007178692030720413, + 0.001624659518711269, + 0.0043821437284350395, + 0.0032029866706579924, + 0.0014458622317761183, + 0.001641030889004469, + 0.0018586438382044435, + 0.0017266152426600456, + 0.0010306491749361157, + 0.00218974775634706, + 0.00374904228374362, + 0.002896655350923538, + 0.0028828720096498728, + 0.004639079328626394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.005256047938019037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062046829611063004 + ], + "opposite_logprob_tokens": [ + -2.812046766281128 + ], + "logprob_reward": -0.062046829611063004, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062046829611063004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03623446822166443, + 0.15973712503910065, + 0.17049291729927063, + 0.12004829943180084, + 0.035267483443021774, + 0.26550379395484924, + 0.08443953096866608, + 0.07247705012559891, + 0.026232389733195305, + 0.02956695295870304 + ], + "credit_map": [ + 0.004572402685880661, + 0.003208027919754386, + 0.0029388596303761005, + 0.0038847781252115965, + 0.004925989080220461, + 0.005269161891192198, + 0.004376530647277832, + 0.007058717776089907, + 0.016664933413267136, + 0.014386773109436035, + 0.0056207794696092606, + 0.011067504063248634, + 0.03639087453484535, + 0.034909602254629135, + 0.019053839147090912, + 0.021642832085490227, + 0.025975381955504417, + 0.010423812083899975, + 0.010202404111623764, + 0.024325648322701454, + 0.01809799298644066, + 0.01672094501554966, + 0.019777530804276466, + 0.04496920853853226, + 0.010544947348535061, + 0.013258322142064571, + 0.007074153982102871, + 0.014191432856023312, + 0.012900970876216888, + 0.011801059357821941, + 0.030086111277341843, + 0.02019129879772663, + 0.0038311798125505447, + 0.0038117198273539543, + 0.003062038216739893, + 0.005666875746101141, + 0.002604837529361248, + 0.0019187652505934238, + 0.007178606931120157, + 0.007193462457507849, + 0.02092079631984234, + 0.021472090855240822, + 0.02490316517651081, + 0.055343519896268845, + 0.014692863449454308, + 0.011922471225261688, + 0.06471849232912064, + 0.0515303835272789, + 0.01375253964215517, + 0.008155351504683495, + 0.005981178488582373, + 0.01305069588124752, + 0.007450539153069258, + 0.006200484465807676, + 0.015889771282672882, + 0.013958970084786415, + 0.006293826270848513, + 0.008725672028958797, + 0.005451522301882505, + 0.008771129883825779, + 0.014036900363862514, + 0.009349022060632706, + 0.008686438202857971, + 0.011162539944052696, + 0.0026414436288177967, + 0.004198699723929167, + 0.0019261494744569063, + 0.0036817272193729877, + 0.004742528777569532, + 0.0036747839767485857, + 0.0026004649698734283, + 0.0027665889356285334, + 0.0024100663140416145, + 0.0027064713649451733, + 0.0013633203925564885, + 0.003087521530687809, + 0.004741370212286711, + 0.0033953567035496235, + 0.004421451594680548, + 0.00744139589369297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.062046829611063004 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281606674194336 + ], + "opposite_logprob_tokens": [ + -2.0178160667419434 + ], + "logprob_reward": -0.14281606674194336, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281606674194336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06971074640750885, + 0.09328831732273102, + 0.09830677509307861, + 0.0466679148375988, + 0.030777467414736748, + 0.13695497810840607, + 0.15391439199447632, + 0.2136894017457962, + 0.05381907895207405, + 0.10287094861268997 + ], + "credit_map": [ + 0.007193780969828367, + 0.006717024836689234, + 0.0075828335247933865, + 0.005747164599597454, + 0.008005838841199875, + 0.007218747865408659, + 0.011204062961041927, + 0.016041278839111328, + 0.01095204520970583, + 0.007460450753569603, + 0.005292728543281555, + 0.008996235206723213, + 0.01789623498916626, + 0.020907163619995117, + 0.010416247881948948, + 0.011367193423211575, + 0.014900065027177334, + 0.006894515827298164, + 0.007942747324705124, + 0.012361814267933369, + 0.008548489771783352, + 0.005840037018060684, + 0.01946149207651615, + 0.022357599809765816, + 0.0041710142977535725, + 0.004316887352615595, + 0.004203091841191053, + 0.005910432431846857, + 0.0043603526428341866, + 0.0026520732790231705, + 0.011457996442914009, + 0.009596060030162334, + 0.003312871791422367, + 0.003272230504080653, + 0.0025445923674851656, + 0.0050866249948740005, + 0.003248736960813403, + 0.0018497309647500515, + 0.0063063595443964005, + 0.005156317260116339, + 0.009518735110759735, + 0.012428178451955318, + 0.010699466802179813, + 0.025889022275805473, + 0.011890305206179619, + 0.009695308282971382, + 0.03251313790678978, + 0.02432081662118435, + 0.027821464464068413, + 0.008705934509634972, + 0.009375913068652153, + 0.018718091771006584, + 0.020791297778487206, + 0.013598864898085594, + 0.019618000835180283, + 0.03528480604290962, + 0.020132774487137794, + 0.02448068931698799, + 0.01715228147804737, + 0.029148275032639503, + 0.03339247405529022, + 0.032747767865657806, + 0.025702575221657753, + 0.030932562425732613, + 0.005003390368074179, + 0.0060707200318574905, + 0.004652573261409998, + 0.010314354673027992, + 0.008547120727598667, + 0.006049982737749815, + 0.006632404867559671, + 0.006548529956489801, + 0.01175983902066946, + 0.00705740787088871, + 0.0068655251525342464, + 0.011964596807956696, + 0.015567502938210964, + 0.012563766911625862, + 0.015405227430164814, + 0.021687081083655357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.14281606674194336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle wide shot, as if shot from a drone's perspective, begins with a slow, steady forward motion, gliding across a tranquil, mirror-like lake that fills the foreground. In the distance, a majestic volcano dominates the frame, its massive mass rising from a lush landscape blanketed by dense, dark green forest. The volcano's summit is partially shrouded in thick, swirling white clouds, while a powerful plume of gray steam and black smoke continuously billows from its crater, creating a stunning contrast against the vast, deep blue sky. As the camera continues its deliberate advance, grand scale gives way to intricate detail; the rugged texture of the dark volcanic rock becomes visible, revealing a complex tapestry of jagged edges and ancient fissures. Occasionally, streams of incandescent lava slowly and deliberately flow down the sides, emitting a fiery orange-red glow that penetrates the dark rock and emits a subtle heat haze, emphasizing the scene's powerful aura of raw, untamed power and sublime natural beauty.", + "seed": 12641, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_8/index_2716_seed_12641_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0212669113650918, + "vlm_logprob": -0.0212669113650918, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.0212669113650918, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03569676727056503 + ], + "opposite_logprob_tokens": [ + -5.160696983337402 + ], + "logprob_reward": -0.03569676727056503, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.03569676727056503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07485257089138031, + 0.11462869495153427, + 0.04156617075204849, + 0.04406658187508583, + 0.08244527131319046, + 0.24562662839889526, + 0.07390456646680832, + 0.040781427174806595, + 0.04823791980743408, + 0.23389014601707458 + ], + "credit_map": [ + 0.00994065124541521, + 0.009135156869888306, + 0.0054001654498279095, + 0.00837971642613411, + 0.012231181375682354, + 0.008855092339217663, + 0.0109470309689641, + 0.009963578544557095, + 0.024251164868474007, + 0.007807358168065548, + 0.0058211772702634335, + 0.006541214417666197, + 0.008111746050417423, + 0.010586073622107506, + 0.01716059446334839, + 0.03434936702251434, + 0.00776614248752594, + 0.003309376770630479, + 0.0022891336120665073, + 0.005226895213127136, + 0.0037652722094208, + 0.0033848967868834734, + 0.005929324310272932, + 0.009895131923258305, + 0.003632280509918928, + 0.003173443954437971, + 0.004391529131680727, + 0.006262857001274824, + 0.004987306892871857, + 0.0038401561323553324, + 0.010019331239163876, + 0.0077596744522452354, + 0.005484817083925009, + 0.004733446519821882, + 0.012379842810332775, + 0.01081217359751463, + 0.007074940949678421, + 0.011531020514667034, + 0.01695912890136242, + 0.013469906523823738, + 0.03522486984729767, + 0.016994204372167587, + 0.03181179240345955, + 0.029367752373218536, + 0.04417983070015907, + 0.01788630709052086, + 0.04073306545615196, + 0.029428817331790924, + 0.003756655612960458, + 0.0037310782354325056, + 0.01470689196139574, + 0.017787059769034386, + 0.003479780163615942, + 0.0030760657973587513, + 0.015413983725011349, + 0.011953049339354038, + 0.003843847895041108, + 0.0033045329619199038, + 0.00688315462321043, + 0.008016597479581833, + 0.0036319796927273273, + 0.00363660859875381, + 0.006128014996647835, + 0.005336692091077566, + 0.0030184676870703697, + 0.003862585173919797, + 0.010698267258703709, + 0.009120103903114796, + 0.002973061753436923, + 0.0029850106220692396, + 0.00886487029492855, + 0.006715552881360054, + 0.011333898641169071, + 0.01560350600630045, + 0.034342844039201736, + 0.03888735920190811, + 0.016471916809678078, + 0.017060771584510803, + 0.06212976574897766, + 0.038060080260038376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.03569676727056503 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822616718709469 + ], + "opposite_logprob_tokens": [ + -4.018226146697998 + ], + "logprob_reward": -0.01822616718709469, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822616718709469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33289000391960144, + 0.23093605041503906, + 0.037903424352407455, + 0.02442014589905739, + 0.045198336243629456, + 0.0696057602763176, + 0.1528913378715515, + 0.029629280790686607, + 0.02306530252099037, + 0.05346030741930008 + ], + "credit_map": [ + 0.04736260697245598, + 0.02647974155843258, + 0.02492225356400013, + 0.03215240314602852, + 0.046562645584344864, + 0.037978995591402054, + 0.052698250859975815, + 0.064733125269413, + 0.047809045761823654, + 0.015859950333833694, + 0.012609036639332771, + 0.021654043346643448, + 0.016566824167966843, + 0.019411733373999596, + 0.03100895881652832, + 0.06601645797491074, + 0.0038952527102082968, + 0.005120342597365379, + 0.003292192006483674, + 0.005235029850155115, + 0.005832736846059561, + 0.002755955094471574, + 0.006882334593683481, + 0.004889583680778742, + 0.0019938310142606497, + 0.0018312057945877314, + 0.0021700889337807894, + 0.005203685257583857, + 0.0020601898431777954, + 0.001383308321237564, + 0.004984375089406967, + 0.00479346327483654, + 0.004532879218459129, + 0.0033964961767196655, + 0.00518685532733798, + 0.006303301081061363, + 0.0049652354791760445, + 0.006118227262049913, + 0.008313721977174282, + 0.006381622515618801, + 0.0038547012954950333, + 0.004544604569673538, + 0.008540024049580097, + 0.014656989835202694, + 0.0056310854852199554, + 0.008046681061387062, + 0.013685247860848904, + 0.010646435432136059, + 0.008650582283735275, + 0.012403919361531734, + 0.023275405168533325, + 0.020071158185601234, + 0.015544729307293892, + 0.011347244493663311, + 0.03363044559955597, + 0.02796785905957222, + 0.0020773604046553373, + 0.001976829022169113, + 0.005536928307265043, + 0.006797002162784338, + 0.0019236321095377207, + 0.002522118389606476, + 0.0048550451174378395, + 0.003940367605537176, + 0.0011655889684334397, + 0.001525915926322341, + 0.00495523028075695, + 0.005086427088826895, + 0.0013972065644338727, + 0.0017451154999434948, + 0.0034528872929513454, + 0.0037369346246123314, + 0.003087091026827693, + 0.003995808772742748, + 0.008655034936964512, + 0.00864956434816122, + 0.0035586904268711805, + 0.004552382044494152, + 0.011426516808569431, + 0.009535220451653004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.01822616718709469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052580637857317924 + ], + "opposite_logprob_tokens": [ + -5.255258083343506 + ], + "logprob_reward": -0.0052580637857317924, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052580637857317924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04701650142669678, + 0.675964891910553, + 0.021904153749346733, + 0.02137690968811512, + 0.022587500512599945, + 0.07767044752836227, + 0.030043719336390495, + 0.025234943255782127, + 0.019350482150912285, + 0.05885044112801552 + ], + "credit_map": [ + 0.005068413447588682, + 0.003909239545464516, + 0.0031836468260735273, + 0.006936710327863693, + 0.005582340527325869, + 0.0038578566163778305, + 0.010682647116482258, + 0.0077956439927220345, + 0.13438652455806732, + 0.0613020583987236, + 0.03257918357849121, + 0.05037077143788338, + 0.05802324786782265, + 0.06100241467356682, + 0.09557661414146423, + 0.18272404372692108, + 0.002052104799076915, + 0.001791229471564293, + 0.0021057925187051296, + 0.004788320045918226, + 0.001618671347387135, + 0.0015936280833557248, + 0.003889127867296338, + 0.004065276123583317, + 0.0017563190776854753, + 0.0022299382835626602, + 0.0017821883084252477, + 0.0028306185267865658, + 0.0024377009831368923, + 0.0019329497590661049, + 0.004883037880063057, + 0.0035241565201431513, + 0.001871799468062818, + 0.0015406538732349873, + 0.0033596421126276255, + 0.0034783033188432455, + 0.001701853354461491, + 0.0022870791144669056, + 0.0047014020383358, + 0.0036467676982283592, + 0.0057893432676792145, + 0.004396809730678797, + 0.008755664341151714, + 0.013207250274717808, + 0.009791744872927666, + 0.008456788025796413, + 0.01641649380326271, + 0.010856343433260918, + 0.0024166707880795, + 0.003937119152396917, + 0.004860402550548315, + 0.004422440659254789, + 0.0031300417613238096, + 0.0027750080917030573, + 0.004500326234847307, + 0.004001707304269075, + 0.0032321475446224213, + 0.0024589768145233393, + 0.004272356163710356, + 0.003670155070722103, + 0.002452239627018571, + 0.002469177357852459, + 0.0035338765010237694, + 0.003146011848002672, + 0.001336907153017819, + 0.0013946974650025368, + 0.0035559351090341806, + 0.004877849947661161, + 0.0017214145045727491, + 0.001449690549634397, + 0.00293730734847486, + 0.0020766803063452244, + 0.002463524928316474, + 0.003284788690507412, + 0.008993308991193771, + 0.009771859273314476, + 0.0037136420141905546, + 0.0042136237025260925, + 0.015129932202398777, + 0.011279757134616375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.0052580637857317924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096094518899918 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096094518899918, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096094518899918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07483059912919998, + 0.08351271599531174, + 0.07048681378364563, + 0.02299591340124607, + 0.03062712773680687, + 0.26976796984672546, + 0.11685837805271149, + 0.09827617555856705, + 0.07248607277870178, + 0.16015824675559998 + ], + "credit_map": [ + 0.01046887505799532, + 0.006347290240228176, + 0.005070246756076813, + 0.005910157226026058, + 0.009030467830598354, + 0.009480077773332596, + 0.011851678602397442, + 0.016671810299158096, + 0.014779657125473022, + 0.00806686282157898, + 0.004868076182901859, + 0.009224512614309788, + 0.006691245827823877, + 0.00786550622433424, + 0.011464763432741165, + 0.020552093163132668, + 0.010775532573461533, + 0.006320675369352102, + 0.0048215254209935665, + 0.007313084788620472, + 0.005736550781875849, + 0.0053548431023955345, + 0.01333644799888134, + 0.016828158870339394, + 0.0018776300130411983, + 0.002618077676743269, + 0.0031175571493804455, + 0.00390880461782217, + 0.0024159264285117388, + 0.0016852065455168486, + 0.004047739785164595, + 0.0033249694388359785, + 0.0029948733281344175, + 0.0023620943538844585, + 0.0037389786448329687, + 0.004546309821307659, + 0.003659304464235902, + 0.0024776833597570658, + 0.006130900699645281, + 0.004716983996331692, + 0.03226063400506973, + 0.017864475026726723, + 0.03276316449046135, + 0.03658125549554825, + 0.04616270959377289, + 0.02167658880352974, + 0.049244947731494904, + 0.03321421518921852, + 0.0072190649807453156, + 0.009019139222800732, + 0.01693785935640335, + 0.017879469320178032, + 0.013051226735115051, + 0.007957037538290024, + 0.02773209661245346, + 0.0170624777674675, + 0.008426434360444546, + 0.008515154011547565, + 0.017785830423235893, + 0.020902568474411964, + 0.008611197583377361, + 0.007145004812628031, + 0.015745103359222412, + 0.011144878342747688, + 0.007010414265096188, + 0.007434163708239794, + 0.012732191011309624, + 0.015136540867388248, + 0.007603807840496302, + 0.005184604786336422, + 0.009550166316330433, + 0.007834183983504772, + 0.0067193033173680305, + 0.008967372588813305, + 0.023064812645316124, + 0.02738688513636589, + 0.010337292216718197, + 0.010931285098195076, + 0.04375436529517174, + 0.028996940702199936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.004096094518899918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305746406316757 + ], + "opposite_logprob_tokens": [ + -3.168057441711426 + ], + "logprob_reward": -0.04305746406316757, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305746406316757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07168187201023102, + 0.03731357306241989, + 0.04980507865548134, + 0.03150070086121559, + 0.06969612091779709, + 0.2065231055021286, + 0.16307206451892853, + 0.09349171072244644, + 0.04557317867875099, + 0.231342613697052 + ], + "credit_map": [ + 0.009907548315823078, + 0.005690963473170996, + 0.004454190377146006, + 0.006370035000145435, + 0.009027807973325253, + 0.010397744365036488, + 0.01131872646510601, + 0.014514856040477753, + 0.005613702815026045, + 0.004547801334410906, + 0.002397331176325679, + 0.005074554588645697, + 0.003961802925914526, + 0.003054728265851736, + 0.005826171021908522, + 0.006837476510554552, + 0.00769333029165864, + 0.0032524981070309877, + 0.0041443826630711555, + 0.007309163920581341, + 0.00379701005294919, + 0.003752441145479679, + 0.009538586251437664, + 0.010317662730813026, + 0.0021601025946438313, + 0.002905680797994137, + 0.0038692993111908436, + 0.006257983855903149, + 0.003994625527411699, + 0.0025230105966329575, + 0.005471108015626669, + 0.004318887833505869, + 0.006966762710362673, + 0.004043158143758774, + 0.010130362585186958, + 0.009163205511868, + 0.0068340059369802475, + 0.004569928161799908, + 0.016482867300510406, + 0.011505825445055962, + 0.022775799036026, + 0.01329630147665739, + 0.02452673949301243, + 0.029621003195643425, + 0.0325997918844223, + 0.018355615437030792, + 0.03916904330253601, + 0.026178792119026184, + 0.014798651449382305, + 0.019593263044953346, + 0.025178316980600357, + 0.023859640583395958, + 0.024086279794573784, + 0.014004652388393879, + 0.022742165252566338, + 0.01880907639861107, + 0.007378949783742428, + 0.006320950575172901, + 0.018061086535453796, + 0.02183632366359234, + 0.006057369522750378, + 0.0075293174013495445, + 0.01543351262807846, + 0.010874190367758274, + 0.004035786259919405, + 0.004051310941576958, + 0.007535293232649565, + 0.010633159428834915, + 0.005298331845551729, + 0.003725099842995405, + 0.005583828315138817, + 0.004710364155471325, + 0.009580268524587154, + 0.013189881108701229, + 0.03314971551299095, + 0.03886253386735916, + 0.014638376422226429, + 0.01579863764345646, + 0.06303280591964722, + 0.043090369552373886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.04305746406316757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle wide shot, as if shot from a drone's perspective, begins with a slow, steady forward motion, gliding across a tranquil, mirror-like lake that fills the foreground. In the distance, a majestic volcano dominates the frame, its massive mass rising from a lush landscape blanketed by dense, dark green forest. The volcano's summit is partially shrouded in thick, swirling white clouds, while a powerful plume of gray steam and black smoke continuously billows from its crater, creating a stunning contrast against the vast, deep blue sky. As the camera continues its deliberate advance, grand scale gives way to intricate detail; the rugged texture of the dark volcanic rock becomes visible, revealing a complex tapestry of jagged edges and ancient fissures. Occasionally, streams of incandescent lava slowly and deliberately flow down the sides, emitting a fiery orange-red glow that penetrates the dark rock and emits a subtle heat haze, emphasizing the scene's powerful aura of raw, untamed power and sublime natural beauty.", + "seed": 12642, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_8/index_2716_seed_12642_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.039514030981808904, + "vlm_logprob": -0.039514030981808904, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.039514030981808904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05401043966412544 + ], + "opposite_logprob_tokens": [ + -3.9290103912353516 + ], + "logprob_reward": -0.05401043966412544, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.05401043966412544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13209843635559082, + 0.28110945224761963, + 0.013101575896143913, + 0.010644882917404175, + 0.0114271380007267, + 0.031397610902786255, + 0.015774531289935112, + 0.04791877046227455, + 0.06650904566049576, + 0.3900185525417328 + ], + "credit_map": [ + 0.024390054866671562, + 0.011296137236058712, + 0.010997750796377659, + 0.00810404121875763, + 0.010901909321546555, + 0.008087897673249245, + 0.02315080538392067, + 0.03516984358429909, + 0.05033525079488754, + 0.020709261298179626, + 0.014681147411465645, + 0.011082129552960396, + 0.008990319445729256, + 0.01854262314736843, + 0.06178778409957886, + 0.09498094022274017, + 0.0012272055028006434, + 0.0011271919356659055, + 0.0010755538241937757, + 0.0017341667553409934, + 0.0016916942549869418, + 0.0011047166772186756, + 0.002276725135743618, + 0.0028643214609473944, + 0.0008206406491808593, + 0.0008624882320873439, + 0.0014303532661870122, + 0.0018210283014923334, + 0.0013651525368914008, + 0.0006735986098647118, + 0.001734651275910437, + 0.0019369696965441108, + 0.0010654791258275509, + 0.0010763583704829216, + 0.0011431440943852067, + 0.0017378279007971287, + 0.0017912668408825994, + 0.0012277072528377175, + 0.0017435167683288455, + 0.0016418375307694077, + 0.002941659651696682, + 0.0033836676739156246, + 0.005337923765182495, + 0.005073690786957741, + 0.0031254286877810955, + 0.0019498929614201188, + 0.004383539315313101, + 0.005201806779950857, + 0.0008108083275146782, + 0.0016169978771358728, + 0.002129551023244858, + 0.0022375709377229214, + 0.00247749756090343, + 0.0014278804883360863, + 0.002592195989564061, + 0.00248203007504344, + 0.005518308840692043, + 0.00839406531304121, + 0.005472343880683184, + 0.0056101297959685326, + 0.006447972729802132, + 0.004576954059302807, + 0.0060726916417479515, + 0.0058263023383915424, + 0.004088451154530048, + 0.006876739207655191, + 0.012081199325621128, + 0.006748269312083721, + 0.014633428305387497, + 0.0060699013993144035, + 0.00806779507547617, + 0.007943259552121162, + 0.02251332812011242, + 0.032126158475875854, + 0.009869523346424103, + 0.015424172393977642, + 0.021432776004076004, + 0.012302211485803127, + 0.1478758156299591, + 0.12847459316253662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.05401043966412544 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870019853115082 + ], + "opposite_logprob_tokens": [ + -3.0487000942230225 + ], + "logprob_reward": -0.04870019853115082, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870019853115082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24248124659061432, + 0.24947576224803925, + 0.028903111815452576, + 0.050397638231515884, + 0.04932786524295807, + 0.034273188561201096, + 0.048078764230012894, + 0.052290547639131546, + 0.06120559945702553, + 0.18356622755527496 + ], + "credit_map": [ + 0.04522312059998512, + 0.020474081858992577, + 0.019586758688092232, + 0.018270790576934814, + 0.01558859646320343, + 0.015267293900251389, + 0.041854288429021835, + 0.06621633470058441, + 0.04357602074742317, + 0.01871202513575554, + 0.01538669504225254, + 0.010367311537265778, + 0.008571877144277096, + 0.015396625734865665, + 0.05435359477996826, + 0.08311164379119873, + 0.0031741487327963114, + 0.0031178933568298817, + 0.0028945747762918472, + 0.0029321019537746906, + 0.004428225103765726, + 0.002318337559700012, + 0.00439473707228899, + 0.005643098149448633, + 0.003409978235140443, + 0.004801323637366295, + 0.004113454837352037, + 0.0071611590683460236, + 0.0077940491028130054, + 0.005624345038086176, + 0.00932290218770504, + 0.008170435205101967, + 0.0038873276207596064, + 0.004338189493864775, + 0.005413598846644163, + 0.008648118935525417, + 0.006963373627513647, + 0.005217771045863628, + 0.008397961035370827, + 0.006461526732891798, + 0.0028030285611748695, + 0.002693910151720047, + 0.0032084165140986443, + 0.004760073032230139, + 0.004296215251088142, + 0.0028067375533282757, + 0.008404062129557133, + 0.005300749558955431, + 0.0033114415127784014, + 0.004998667165637016, + 0.006375766824930906, + 0.0049700564704835415, + 0.00646613072603941, + 0.0049424320459365845, + 0.00868448056280613, + 0.008329791948199272, + 0.006175375077873468, + 0.010675814934074879, + 0.004402104765176773, + 0.005371416453272104, + 0.005148962140083313, + 0.0037370186764746904, + 0.005645030178129673, + 0.01113483589142561, + 0.0036839903332293034, + 0.006651270668953657, + 0.013285448774695396, + 0.008740954101085663, + 0.00847665499895811, + 0.003593944711610675, + 0.007747294846922159, + 0.009026046842336655, + 0.017912477254867554, + 0.020975708961486816, + 0.008965269662439823, + 0.008808547630906105, + 0.014345317147672176, + 0.007770211901515722, + 0.053738705813884735, + 0.05104999989271164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.04870019853115082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017201402224600315 + ], + "opposite_logprob_tokens": [ + -6.376719951629639 + ], + "logprob_reward": -0.0017201402224600315, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017201402224600315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053160760551691055, + 0.09652109444141388, + 0.026140877977013588, + 0.008589040488004684, + 0.01392518449574709, + 0.05122101306915283, + 0.01595202460885048, + 0.034095004200935364, + 0.05955032259225845, + 0.6408447623252869 + ], + "credit_map": [ + 0.0065125636756420135, + 0.0055493637919425964, + 0.005044078454375267, + 0.005910986103117466, + 0.00679921405389905, + 0.0038130774628371, + 0.00935716088861227, + 0.010174315422773361, + 0.01715789921581745, + 0.007168274372816086, + 0.004917450714856386, + 0.004283395130187273, + 0.003996449057012796, + 0.006145699881017208, + 0.02048259973526001, + 0.03236931934952736, + 0.003037628484889865, + 0.001788030145689845, + 0.0017015046905726194, + 0.0022364980541169643, + 0.0024086488410830498, + 0.002311947988346219, + 0.005045823287218809, + 0.0076107969507575035, + 0.0007512898882851005, + 0.0007802829495631158, + 0.0007163357804529369, + 0.0011360901407897472, + 0.001220006961375475, + 0.0006985634099692106, + 0.0017467050347477198, + 0.0015397662064060569, + 0.0010399111779406667, + 0.0012287714052945375, + 0.0014427918940782547, + 0.0025960297789424658, + 0.002606423804536462, + 0.001209668000228703, + 0.0021512326784431934, + 0.001650356687605381, + 0.0033179984893649817, + 0.004626052919775248, + 0.005875944159924984, + 0.0056040408089756966, + 0.009459232911467552, + 0.005221573170274496, + 0.009273063391447067, + 0.007843106985092163, + 0.0008772353176027536, + 0.001901010749861598, + 0.002529341960325837, + 0.0017978857504203916, + 0.002177129965275526, + 0.0012687025591731071, + 0.003133482299745083, + 0.0022672356572002172, + 0.00269529246725142, + 0.005827191285789013, + 0.005326845217496157, + 0.004587430041283369, + 0.004687185399234295, + 0.002138998592272401, + 0.004402690567076206, + 0.004429370164871216, + 0.0032825402449816465, + 0.006375690922141075, + 0.00477825291454792, + 0.008058374747633934, + 0.016070442274212837, + 0.0058142258785665035, + 0.007466960232704878, + 0.007703833747655153, + 0.038260288536548615, + 0.08401697874069214, + 0.015453796833753586, + 0.02446846291422844, + 0.039906036108732224, + 0.02044624462723732, + 0.2252698540687561, + 0.19302304089069366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.0017201402224600315 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096094518899918 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096094518899918, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096094518899918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055422089993953705, + 0.18937036395072937, + 0.026775117963552475, + 0.00612466037273407, + 0.005027545150369406, + 0.033312395215034485, + 0.009616302326321602, + 0.02243644744157791, + 0.03752633184194565, + 0.6143887639045715 + ], + "credit_map": [ + 0.010973332449793816, + 0.005245269741863012, + 0.0035061584785580635, + 0.0024306466802954674, + 0.003929573576897383, + 0.0032807039096951485, + 0.010272873565554619, + 0.015783535316586494, + 0.033969487994909286, + 0.014036750420928001, + 0.010049345903098583, + 0.007226655725389719, + 0.006710503250360489, + 0.011828255839645863, + 0.041490789502859116, + 0.06405859440565109, + 0.003283795900642872, + 0.001537048607133329, + 0.0014918922679498792, + 0.001952151651494205, + 0.0019277905812487006, + 0.0016628061421215534, + 0.006103551480919123, + 0.008816083893179893, + 0.0004082348314113915, + 0.0004491799627430737, + 0.0005698805907741189, + 0.0009703858522698283, + 0.0007439666660502553, + 0.00045139159192331135, + 0.0012346557341516018, + 0.0012969657545909286, + 0.0003855074755847454, + 0.0006096299039199948, + 0.0005516730598174036, + 0.000718321418389678, + 0.0007181147811934352, + 0.00035219232086092234, + 0.0009487440693192184, + 0.0007433622376993299, + 0.0020600329153239727, + 0.0030907499603927135, + 0.0031687787268310785, + 0.004324594512581825, + 0.005188381765037775, + 0.0029509123414754868, + 0.006766293663531542, + 0.005762654822319746, + 0.0007467052200809121, + 0.0008155368850566447, + 0.0009351299959234893, + 0.0010173877235502005, + 0.0014405719703063369, + 0.0009746852447278798, + 0.001555248978547752, + 0.002131036948412657, + 0.001327856327407062, + 0.002064354717731476, + 0.004491155035793781, + 0.004055439960211515, + 0.003038478083908558, + 0.0017876379424706101, + 0.003569943131878972, + 0.0021015831734985113, + 0.002111243549734354, + 0.004053117707371712, + 0.003715052269399166, + 0.003114851890131831, + 0.00983696524053812, + 0.0034697011578828096, + 0.00530938571318984, + 0.005916016176342964, + 0.03543395921587944, + 0.05949477106332779, + 0.015383753925561905, + 0.024707181379199028, + 0.033884141594171524, + 0.019852686673402786, + 0.22696612775325775, + 0.19866609573364258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.004096094518899918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890432819724083 + ], + "opposite_logprob_tokens": [ + -2.464043378829956 + ], + "logprob_reward": -0.0890432819724083, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890432819724083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05018625035881996, + 0.35158368945121765, + 0.0343339703977108, + 0.014307214878499508, + 0.01246540155261755, + 0.10259442031383514, + 0.026218144223093987, + 0.04567798972129822, + 0.0906677395105362, + 0.2719652056694031 + ], + "credit_map": [ + 0.007521539460867643, + 0.004630655515938997, + 0.0063401274383068085, + 0.00469773868098855, + 0.006161903962492943, + 0.0035627668257802725, + 0.0073992847464978695, + 0.00987223256379366, + 0.06274603307247162, + 0.02615399844944477, + 0.01891261897981167, + 0.013265418820083141, + 0.013035268522799015, + 0.02160480059683323, + 0.07725327461957932, + 0.11861225217580795, + 0.003485012799501419, + 0.002343878149986267, + 0.0024513399694114923, + 0.0032126596197485924, + 0.004236734937876463, + 0.0030177240259945393, + 0.006440871395170689, + 0.009145750664174557, + 0.0009398071561008692, + 0.0010331887751817703, + 0.0011995547683909535, + 0.0025780911091715097, + 0.0017328511457890272, + 0.0008722697966732085, + 0.002974987030029297, + 0.0029764650389552116, + 0.000908609654288739, + 0.0013559333747252822, + 0.0011644675396382809, + 0.002202259609475732, + 0.001864338992163539, + 0.001094707869924605, + 0.002198031172156334, + 0.0016770531656220555, + 0.0054568275809288025, + 0.00847664475440979, + 0.013307466171681881, + 0.012146830558776855, + 0.01617267355322838, + 0.010217436589300632, + 0.0213716309517622, + 0.01544491108506918, + 0.001806116895750165, + 0.0028062842320650816, + 0.004023529589176178, + 0.003000101540237665, + 0.0032242683228105307, + 0.0020798002369701862, + 0.0038004389498382807, + 0.0054776049219071865, + 0.0023803538642823696, + 0.004627881105989218, + 0.009356999769806862, + 0.006296148058027029, + 0.005681934300810099, + 0.0034104040823876858, + 0.007191838696599007, + 0.006732431706041098, + 0.004876857623457909, + 0.010504168458282948, + 0.01122624333947897, + 0.012599513866007328, + 0.018057746812701225, + 0.009945153258740902, + 0.012401380576193333, + 0.011056676506996155, + 0.019011182710528374, + 0.015192190185189247, + 0.012173116207122803, + 0.012705976143479347, + 0.014893591403961182, + 0.009329769760370255, + 0.09852300584316254, + 0.09013639390468597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.0890432819724083 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle wide shot, as if shot from a drone's perspective, begins with a slow, steady forward motion, gliding across a tranquil, mirror-like lake that fills the foreground. In the distance, a majestic volcano dominates the frame, its massive mass rising from a lush landscape blanketed by dense, dark green forest. The volcano's summit is partially shrouded in thick, swirling white clouds, while a powerful plume of gray steam and black smoke continuously billows from its crater, creating a stunning contrast against the vast, deep blue sky. As the camera continues its deliberate advance, grand scale gives way to intricate detail; the rugged texture of the dark volcanic rock becomes visible, revealing a complex tapestry of jagged edges and ancient fissures. Occasionally, streams of incandescent lava slowly and deliberately flow down the sides, emitting a fiery orange-red glow that penetrates the dark rock and emits a subtle heat haze, emphasizing the scene's powerful aura of raw, untamed power and sublime natural beauty.", + "seed": 12643, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_8/index_2716_seed_12643_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04325810959562659, + "vlm_logprob": -0.04325810959562659, + "vlm_yes_no_margin": 4.125000023841858, + "vlm_token_credit": -0.04325810959562659, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03427641838788986 + ], + "opposite_logprob_tokens": [ + -4.534276485443115 + ], + "logprob_reward": -0.03427641838788986, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.03427641838788986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11007066816091537, + 0.08312761783599854, + 0.04668446257710457, + 0.0877232626080513, + 0.07483942806720734, + 0.06962317228317261, + 0.0678643062710762, + 0.1529524028301239, + 0.048770464956760406, + 0.25834420323371887 + ], + "credit_map": [ + 0.011835962533950806, + 0.011998365633189678, + 0.009829523041844368, + 0.01492609828710556, + 0.016487490385770798, + 0.017094390466809273, + 0.016461564227938652, + 0.011437270790338516, + 0.017101548612117767, + 0.009149444289505482, + 0.004373634699732065, + 0.007263080216944218, + 0.00657100323587656, + 0.007044903002679348, + 0.014602679759263992, + 0.01702132448554039, + 0.00550354178994894, + 0.004556987900286913, + 0.0037703714333474636, + 0.0029717269353568554, + 0.0034028510563075542, + 0.0064632040448486805, + 0.010719391517341137, + 0.009296388365328312, + 0.009389273822307587, + 0.007102672941982746, + 0.007940767332911491, + 0.006792103871703148, + 0.008825932629406452, + 0.016670918092131615, + 0.01651550643146038, + 0.01448608748614788, + 0.007400416769087315, + 0.005640543065965176, + 0.011138354428112507, + 0.010437841527163982, + 0.005868326872587204, + 0.012810696847736835, + 0.010131778195500374, + 0.011411472223699093, + 0.008719980716705322, + 0.011593728326261044, + 0.004951164126396179, + 0.005603722296655178, + 0.011758190579712391, + 0.010632866993546486, + 0.009290846064686775, + 0.0070726703852415085, + 0.007900048978626728, + 0.004746024031192064, + 0.00799996592104435, + 0.012877020053565502, + 0.0044980705715715885, + 0.009941850788891315, + 0.008419327437877655, + 0.011482001282274723, + 0.010531811974942684, + 0.008177201263606548, + 0.010897016152739525, + 0.035773955285549164, + 0.00953185185790062, + 0.013035673648118973, + 0.03340531140565872, + 0.03159957006573677, + 0.0026248879730701447, + 0.003628550097346306, + 0.006853473372757435, + 0.010522456839680672, + 0.0032302176114171743, + 0.004548443481326103, + 0.00964308064430952, + 0.007719356566667557, + 0.02612130530178547, + 0.0226435624063015, + 0.044988781213760376, + 0.04472672939300537, + 0.01274366769939661, + 0.019705263897776604, + 0.017353594303131104, + 0.07006131857633591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.03427641838788986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02984113246202469 + ], + "opposite_logprob_tokens": [ + -3.529841184616089 + ], + "logprob_reward": -0.02984113246202469, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02984113246202469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1563713699579239, + 0.12999723851680756, + 0.0317806638777256, + 0.045643966645002365, + 0.04844389483332634, + 0.056957729160785675, + 0.12170693278312683, + 0.033254291862249374, + 0.05306581035256386, + 0.3227781057357788 + ], + "credit_map": [ + 0.017145775258541107, + 0.015620690770447254, + 0.013666578568518162, + 0.017368212342262268, + 0.012968198396265507, + 0.013102655299007893, + 0.03675203397870064, + 0.029747221618890762, + 0.018698781728744507, + 0.014703815802931786, + 0.014739740639925003, + 0.01325785182416439, + 0.009905442595481873, + 0.015445630997419357, + 0.023849330842494965, + 0.019396651536226273, + 0.003937254659831524, + 0.004136895760893822, + 0.004201574716717005, + 0.004720164928585291, + 0.003214424941688776, + 0.004126398824155331, + 0.0038201920688152313, + 0.003623758442699909, + 0.0049224672839045525, + 0.004237104672938585, + 0.004524971358478069, + 0.005214357282966375, + 0.004697556607425213, + 0.008562322705984116, + 0.007627388928085566, + 0.005857802461832762, + 0.005074196495115757, + 0.006438867654651403, + 0.004226019140332937, + 0.004606037866324186, + 0.005904881749302149, + 0.009301774203777313, + 0.00709364702925086, + 0.005798474419862032, + 0.006037297658622265, + 0.0063622258603572845, + 0.004943772219121456, + 0.004275610204786062, + 0.008179590106010437, + 0.0076870848424732685, + 0.011444473639130592, + 0.008027676492929459, + 0.010509040206670761, + 0.01111317053437233, + 0.015532727353274822, + 0.02098941057920456, + 0.012844221666455269, + 0.018611770123243332, + 0.016173508018255234, + 0.015933088958263397, + 0.0027492207009345293, + 0.0026263410691171885, + 0.003510484006255865, + 0.005720444023609161, + 0.0022997334599494934, + 0.004249176941812038, + 0.005501564592123032, + 0.006597326602786779, + 0.002640931401401758, + 0.0024090437218546867, + 0.006836169399321079, + 0.010281281545758247, + 0.004688112996518612, + 0.004781427793204784, + 0.012591823935508728, + 0.0088370181620121, + 0.019912315532565117, + 0.01276494562625885, + 0.023297108709812164, + 0.09896309673786163, + 0.0075604841113090515, + 0.01766177825629711, + 0.019240956753492355, + 0.12337739765644073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.02984113246202469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099537618458271 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099537618458271, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099537618458271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06117777153849602, + 0.10318361222743988, + 0.13307562470436096, + 0.13511520624160767, + 0.10023867338895798, + 0.07870395481586456, + 0.09296900779008865, + 0.13596636056900024, + 0.06183871254324913, + 0.0977310836315155 + ], + "credit_map": [ + 0.008286754600703716, + 0.006879517808556557, + 0.0037609401624649763, + 0.006899903062731028, + 0.005362809170037508, + 0.006430777255445719, + 0.013655653223395348, + 0.009901408106088638, + 0.01718105562031269, + 0.009922662749886513, + 0.007383349351584911, + 0.011831196956336498, + 0.00924069993197918, + 0.013596986420452595, + 0.01671886257827282, + 0.017308803275227547, + 0.017571348696947098, + 0.011956787668168545, + 0.010327021591365337, + 0.011173424310982227, + 0.015735294669866562, + 0.017053227871656418, + 0.02800695225596428, + 0.021251581609249115, + 0.015748588368296623, + 0.011433389969170094, + 0.010436635464429855, + 0.009984686970710754, + 0.014063537120819092, + 0.035512786358594894, + 0.021687714383006096, + 0.016247868537902832, + 0.008957733400166035, + 0.010830383747816086, + 0.008098592981696129, + 0.010871094651520252, + 0.007039104588329792, + 0.019209453836083412, + 0.0200112946331501, + 0.015221018344163895, + 0.007090070750564337, + 0.00987348984926939, + 0.007693316321820021, + 0.008763039484620094, + 0.0065513537265360355, + 0.013221328146755695, + 0.01382916234433651, + 0.011682195588946342, + 0.009369309060275555, + 0.00682316767051816, + 0.011800273321568966, + 0.01344506349414587, + 0.0076812985353171825, + 0.01631595939397812, + 0.013404106721282005, + 0.014129829593002796, + 0.014731164090335369, + 0.015644624829292297, + 0.012784160673618317, + 0.02440609224140644, + 0.013920697383582592, + 0.013257686980068684, + 0.01898101717233658, + 0.022240914404392242, + 0.0036239358596503735, + 0.0029576842207461596, + 0.007722136564552784, + 0.01073357556015253, + 0.00442537060007453, + 0.006318061146885157, + 0.015148584730923176, + 0.010909365490078926, + 0.006755348294973373, + 0.01160207949578762, + 0.01631493866443634, + 0.017216695472598076, + 0.01090114377439022, + 0.008670555427670479, + 0.009494324214756489, + 0.016775991767644882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.004099537618458271 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255336407572031 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255336407572031, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255336407572031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048308104276657104, + 0.05845307931303978, + 0.018772859126329422, + 0.03306053578853607, + 0.03570714220404625, + 0.027578873559832573, + 0.06319914013147354, + 0.12152764201164246, + 0.051424700766801834, + 0.5419679284095764 + ], + "credit_map": [ + 0.004933846183121204, + 0.004510390572249889, + 0.0053658075630664825, + 0.00577895250171423, + 0.006172248162329197, + 0.006268293596804142, + 0.009272882714867592, + 0.006005682051181793, + 0.01264748815447092, + 0.0068176812492311, + 0.0033593999687582254, + 0.003458345076069236, + 0.004256980028003454, + 0.005618862342089415, + 0.009026640094816685, + 0.013267682865262032, + 0.002414006507024169, + 0.0019884558860212564, + 0.0019007952651008964, + 0.002080404432490468, + 0.002057968871667981, + 0.0037090943660587072, + 0.002576071536168456, + 0.002046062145382166, + 0.0033629867248237133, + 0.00257027056068182, + 0.003527059219777584, + 0.0038662797305732965, + 0.0021108093205839396, + 0.005631553940474987, + 0.007203410379588604, + 0.0047881645150482655, + 0.002724785590544343, + 0.0038890682626515627, + 0.002806856529787183, + 0.003588140243664384, + 0.002858887193724513, + 0.007155733648687601, + 0.007254701107740402, + 0.005428970325738192, + 0.002015417907387018, + 0.001758942729793489, + 0.00263134203851223, + 0.004431325942277908, + 0.0019088475964963436, + 0.0032183669973164797, + 0.006352398544549942, + 0.005262231919914484, + 0.010367752984166145, + 0.0042615896090865135, + 0.005043637473136187, + 0.005651008803397417, + 0.004187993239611387, + 0.004164398647844791, + 0.011538032442331314, + 0.017984725534915924, + 0.007072080858051777, + 0.005481587257236242, + 0.007696312852203846, + 0.027678750455379486, + 0.006952160969376564, + 0.011630094610154629, + 0.02947411499917507, + 0.025542544201016426, + 0.0027817082591354847, + 0.0028598527424037457, + 0.007152019068598747, + 0.010256881825625896, + 0.0037555459421128035, + 0.005928392522037029, + 0.010706130415201187, + 0.007984166964888573, + 0.03236948698759079, + 0.02233739010989666, + 0.03905683010816574, + 0.17106707394123077, + 0.01220325194299221, + 0.028633149340748787, + 0.03177750110626221, + 0.2045232504606247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.005255336407572031 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, resembling a drone's perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow, steady forward motion gliding across a mirror-like lake in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large volcano visible in the distance with a summit partially covered by thick, swirling white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are streams of incandescent lava flowing down the volcano's sides, emitting an orange-red glow and heat haze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video reveal the rugged texture of the dark volcanic rock, including jagged edges and ancient fissures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428181231021881 + ], + "opposite_logprob_tokens": [ + -2.0178182125091553 + ], + "logprob_reward": -0.1428181231021881, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1428181231021881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022984357550740242, + 0.03241444006562233, + 0.034515298902988434, + 0.019267771393060684, + 0.03035956807434559, + 0.018244659528136253, + 0.05674777925014496, + 0.0676097646355629, + 0.032625362277030945, + 0.6852310299873352 + ], + "credit_map": [ + 0.0025796773843467236, + 0.0038418425247073174, + 0.0021520399022847414, + 0.0020155140664428473, + 0.002415039809420705, + 0.0025995050091296434, + 0.003376125590875745, + 0.004004613030701876, + 0.005792577285319567, + 0.003521319478750229, + 0.00196297955699265, + 0.0025041981134563684, + 0.0016876091249287128, + 0.0026363483630120754, + 0.007744967006146908, + 0.006564442999660969, + 0.004549801349639893, + 0.0034718457609415054, + 0.002448643557727337, + 0.0031607861164957285, + 0.00436500133946538, + 0.005043296609073877, + 0.006846790201961994, + 0.004629133269190788, + 0.001890872372314334, + 0.0015026697656139731, + 0.001969157485291362, + 0.0027501797303557396, + 0.0012857664842158556, + 0.002366663422435522, + 0.003796517848968506, + 0.003705944400280714, + 0.0027075083926320076, + 0.0030630817636847496, + 0.004004558082669973, + 0.0032212291844189167, + 0.002382757142186165, + 0.006415374577045441, + 0.00413168128579855, + 0.004433377180248499, + 0.0019658030942082405, + 0.0025656060315668583, + 0.0022918207105249166, + 0.0019991786684840918, + 0.0020738968160003424, + 0.0019166605779901147, + 0.0030399051029235125, + 0.0023917891085147858, + 0.008289159275591373, + 0.0035905237309634686, + 0.004680568352341652, + 0.009118583053350449, + 0.003078657202422619, + 0.003741798922419548, + 0.009226592257618904, + 0.015021898783743382, + 0.0041046603582799435, + 0.0032148705795407295, + 0.004954650066792965, + 0.014796256087720394, + 0.003983138594776392, + 0.006331103853881359, + 0.016572505235671997, + 0.013652583584189415, + 0.001771103241480887, + 0.0017557189567014575, + 0.004138296004384756, + 0.0063336798921227455, + 0.0028578334022313356, + 0.004173944238573313, + 0.006520216353237629, + 0.005074569955468178, + 0.04245352745056152, + 0.027983618900179863, + 0.04549097269773483, + 0.2184225469827652, + 0.013234664686024189, + 0.035475924611091614, + 0.04033002257347107, + 0.26183971762657166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.1428181231021881 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_9/index_tensor([7300, 7300])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_9/index_tensor([7300, 7300])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..507252fac2ac379f7b5c77a187c36ce046edd997 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_9/index_tensor([7300, 7300])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, steady shot, shot at eye level, captures a tranquil, sun-drenched scene. An elderly man, his face etched with soft wrinkles, his silver-white hair gleaming in the light, methodically mows a vast, vibrant green lawn. Wearing a faded light blue work shirt with sleeves rolled up to the elbows and simple brown trousers, he maintains a slightly bent posture as he steadily and deliberately pushes a classic, dark green manual reel mower. With each push, the mower's blades rotate with a quiet, mechanical whir, neatly mowing the emerald green grass and leaving a perfectly straight, slightly darkened path in their wake. In the background, a row of charming storybook cottages sit beneath a clear sky, their white stone walls adorned with climbing roses and their window sills overflowing with flower boxes. Their most striking feature is their thick, intricately tiered roofs, which glow a warm golden-brown in the afternoon sun. The warm light bathes the entire landscape, casting long, soft shadows and creating an idyllic atmosphere of peace and eternal, quiet contentment.", + "seed": 32260, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_9/index_7300_seed_32260_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1311264995019883, + "vlm_logprob": -0.1311264995019883, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.1311264995019883, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4729955792427063 + ], + "opposite_logprob_tokens": [ + -1.8479955196380615 + ], + "logprob_reward": -0.4729955792427063, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.4729955792427063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14448125660419464, + 0.024206029251217842, + 0.27931442856788635, + 0.027038952335715294, + 0.09569850564002991, + 0.04315472021698952, + 0.11556965857744217, + 0.1055147647857666, + 0.07038724422454834, + 0.09463442116975784 + ], + "credit_map": [ + 0.012320918962359428, + 0.012752066366374493, + 0.03314671292901039, + 0.026706689968705177, + 0.016528915613889694, + 0.01171030942350626, + 0.014606972225010395, + 0.016708681359887123, + 0.0014282695483416319, + 0.002052405383437872, + 0.005128604359924793, + 0.004235307686030865, + 0.0025084035005420446, + 0.002511853352189064, + 0.002886211732402444, + 0.003454973455518484, + 0.029087906703352928, + 0.021465422585606575, + 0.03306802734732628, + 0.04199778288602829, + 0.029912307858467102, + 0.044166199862957, + 0.04634896293282509, + 0.03326784819364548, + 0.0017796881729736924, + 0.0015038982965052128, + 0.005043673794716597, + 0.0037841673474758863, + 0.0032855833414942026, + 0.003128140000626445, + 0.004466784652322531, + 0.004047017078846693, + 0.007050890475511551, + 0.008932250551879406, + 0.010698569938540459, + 0.017140060663223267, + 0.009660064242780209, + 0.006635937839746475, + 0.019216055050492287, + 0.016364680603146553, + 0.00320914713665843, + 0.003947730641812086, + 0.0052254898473620415, + 0.005109698511660099, + 0.004536820109933615, + 0.0081154964864254, + 0.007748321630060673, + 0.005262019578367472, + 0.007629474624991417, + 0.006065089255571365, + 0.02393960952758789, + 0.014936760067939758, + 0.02124932035803795, + 0.00962248258292675, + 0.018981680274009705, + 0.013145240023732185, + 0.0067260125651955605, + 0.007625192403793335, + 0.02034049853682518, + 0.01207211334258318, + 0.017389671877026558, + 0.008653673343360424, + 0.018237730488181114, + 0.014469870366156101, + 0.00501394784078002, + 0.0046148221008479595, + 0.007546278648078442, + 0.00999151449650526, + 0.01247195340692997, + 0.005229766946285963, + 0.012017610482871532, + 0.013501355424523354, + 0.005764092784374952, + 0.004180127754807472, + 0.029712041839957237, + 0.009694854728877544, + 0.007246456108987331, + 0.005795286502689123, + 0.004564015660434961, + 0.02767755091190338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.4729955792427063 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036860644817352 + ], + "opposite_logprob_tokens": [ + -1.91036856174469 + ], + "logprob_reward": -0.16036860644817352, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036860644817352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2448105663061142, + 0.030502427369356155, + 0.214995339512825, + 0.025856900960206985, + 0.10176197439432144, + 0.0346301831305027, + 0.1251971423625946, + 0.06917143613100052, + 0.06447286158800125, + 0.08860118687152863 + ], + "credit_map": [ + 0.012607753276824951, + 0.01424435805529356, + 0.040354520082473755, + 0.06553834676742554, + 0.013993507251143456, + 0.024118831381201744, + 0.03417382389307022, + 0.03977940231561661, + 0.0023480625823140144, + 0.002051214687526226, + 0.003648126730695367, + 0.004348930902779102, + 0.003996668849140406, + 0.0033206853549927473, + 0.007076593581587076, + 0.003712147008627653, + 0.02065308764576912, + 0.016453607007861137, + 0.02482311986386776, + 0.03293945640325546, + 0.026553167030215263, + 0.03585698828101158, + 0.03292260691523552, + 0.024793317541480064, + 0.002063895110040903, + 0.002170540625229478, + 0.003084520110860467, + 0.003914056345820427, + 0.002524847164750099, + 0.004966284614056349, + 0.0037242069374769926, + 0.0034085512161254883, + 0.007082505617290735, + 0.004389765206724405, + 0.022191012278199196, + 0.014863932505249977, + 0.011406105943024158, + 0.007340880576521158, + 0.022961845621466637, + 0.011525923386216164, + 0.0024547846987843513, + 0.0033724054228514433, + 0.005803970154374838, + 0.007015120703727007, + 0.002563964808359742, + 0.00523053202778101, + 0.003851996036246419, + 0.004337407648563385, + 0.007656685542315245, + 0.005482559558004141, + 0.027992412447929382, + 0.016139278188347816, + 0.02302233688533306, + 0.00986845139414072, + 0.02006063051521778, + 0.014974790625274181, + 0.004013172350823879, + 0.003476331476122141, + 0.014109047129750252, + 0.00847782101482153, + 0.01342261116951704, + 0.004923350177705288, + 0.011707544326782227, + 0.009041554294526577, + 0.004483568482100964, + 0.005493897944688797, + 0.007266743574291468, + 0.00940152071416378, + 0.010057221166789532, + 0.0067458064295351505, + 0.01319969817996025, + 0.00782440509647131, + 0.0037510059773921967, + 0.0024835322983562946, + 0.030299365520477295, + 0.010665394365787506, + 0.005884996615350246, + 0.003877851413562894, + 0.004494218155741692, + 0.02714482694864273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.16036860644817352 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260554142296314 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260554142296314, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260554142296314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1697172224521637, + 0.042937129735946655, + 0.1082606092095375, + 0.02322869561612606, + 0.1273982673883438, + 0.028557702898979187, + 0.20201200246810913, + 0.09905959665775299, + 0.1264415681362152, + 0.07238715142011642 + ], + "credit_map": [ + 0.012129283510148525, + 0.013754229992628098, + 0.028351502493023872, + 0.03754255175590515, + 0.013720978051424026, + 0.024677690118551254, + 0.01839049905538559, + 0.021150484681129456, + 0.003245403291657567, + 0.003021947108209133, + 0.00539852399379015, + 0.006247145123779774, + 0.004279253073036671, + 0.0064690737053751945, + 0.009644799865782261, + 0.004630983807146549, + 0.00977222342044115, + 0.00875777006149292, + 0.013128301128745079, + 0.015316630713641644, + 0.011777241714298725, + 0.01740681193768978, + 0.018453335389494896, + 0.013648299500346184, + 0.0019953777082264423, + 0.0020452600438147783, + 0.0029204583261162043, + 0.0035939598456025124, + 0.0024056320544332266, + 0.004522363189607859, + 0.00313210510648787, + 0.002613537944853306, + 0.009141313843429089, + 0.007448081858456135, + 0.026267269626259804, + 0.020108435302972794, + 0.013442804105579853, + 0.008918298408389091, + 0.027629196643829346, + 0.014442871324717999, + 0.0017122675199061632, + 0.0017582401633262634, + 0.005255528725683689, + 0.003942080773413181, + 0.004817438777536154, + 0.0028219828382134438, + 0.004618398379534483, + 0.003631766652688384, + 0.014073938131332397, + 0.010631684213876724, + 0.04167193919420242, + 0.028744040057063103, + 0.0313541516661644, + 0.01836434006690979, + 0.029713695868849754, + 0.02745821326971054, + 0.007749210577458143, + 0.0073517076671123505, + 0.014350752346217632, + 0.015471765771508217, + 0.022133538499474525, + 0.00930042564868927, + 0.013636281713843346, + 0.00906591210514307, + 0.00734957680106163, + 0.010305359959602356, + 0.014145025052130222, + 0.018174849450588226, + 0.020592542365193367, + 0.01052041631191969, + 0.02524305321276188, + 0.020110754296183586, + 0.0047325859777629375, + 0.003166934009641409, + 0.021279899403452873, + 0.008885746821761131, + 0.004940786398947239, + 0.004707466345280409, + 0.005220154765993357, + 0.01945357397198677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.005260554142296314 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820919966325164 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.002820919966325164, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820919966325164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1545717716217041, + 0.05455461144447327, + 0.39069241285324097, + 0.051493559032678604, + 0.07273146510124207, + 0.020527686923742294, + 0.03950055316090584, + 0.035697273910045624, + 0.061626043170690536, + 0.11860455572605133 + ], + "credit_map": [ + 0.012166636064648628, + 0.01180244144052267, + 0.022386083379387856, + 0.02819603867828846, + 0.013831297867000103, + 0.023331603035330772, + 0.023760920390486717, + 0.019096771255135536, + 0.005239832680672407, + 0.0056272828951478004, + 0.0067126513458788395, + 0.006397436372935772, + 0.004743156488984823, + 0.009040613658726215, + 0.0111535070464015, + 0.005640135612338781, + 0.03895688056945801, + 0.029852043837308884, + 0.04886908084154129, + 0.060566484928131104, + 0.041898611932992935, + 0.06317435204982758, + 0.06087518855929375, + 0.046499818563461304, + 0.003697959939017892, + 0.003895549336448312, + 0.00992901623249054, + 0.007745199371129274, + 0.00403223279863596, + 0.007594119757413864, + 0.00893922708928585, + 0.005660261958837509, + 0.006469583138823509, + 0.005691646598279476, + 0.012706635519862175, + 0.013342352584004402, + 0.00645682355388999, + 0.007400744128972292, + 0.012084213085472584, + 0.008579478599131107, + 0.001527733402326703, + 0.0012991168769076467, + 0.002855348400771618, + 0.002259156433865428, + 0.003095872700214386, + 0.003110339865088463, + 0.0038998418021947145, + 0.0024802808184176683, + 0.005561972036957741, + 0.0037820138968527317, + 0.006039351224899292, + 0.00671294005587697, + 0.004378844518214464, + 0.004134319256991148, + 0.0043798089027404785, + 0.004511307459324598, + 0.004022971261292696, + 0.003974082414060831, + 0.005304948426783085, + 0.007134095765650272, + 0.00273897685110569, + 0.0039017428644001484, + 0.0029829207342118025, + 0.005637543275952339, + 0.003934523556381464, + 0.004130203742533922, + 0.013710708357393742, + 0.00834037084132433, + 0.00932647380977869, + 0.004989905282855034, + 0.006919059436768293, + 0.010274806059896946, + 0.007141479291021824, + 0.004669313319027424, + 0.03804115578532219, + 0.01502193883061409, + 0.007863006554543972, + 0.006732996087521315, + 0.0067993039265275, + 0.032335374504327774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.002820919966325164 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186837710440159 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.014186837710440159, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186837710440159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10367568582296371, + 0.020050156861543655, + 0.1546543538570404, + 0.03005872666835785, + 0.20246951282024384, + 0.007478306535631418, + 0.06373640894889832, + 0.03895020857453346, + 0.015294323675334454, + 0.36363232135772705 + ], + "credit_map": [ + 0.011049589142203331, + 0.009559407830238342, + 0.01316390372812748, + 0.013668320141732693, + 0.012337759137153625, + 0.018600592389702797, + 0.01395406387746334, + 0.011342047713696957, + 0.0017458571819588542, + 0.001776961493305862, + 0.0029175945091992617, + 0.0031549339182674885, + 0.0014298732858151197, + 0.003146058414131403, + 0.00374671071767807, + 0.0021321678068488836, + 0.014931951649487019, + 0.012381860055029392, + 0.021099500358104706, + 0.02336750365793705, + 0.016758780926465988, + 0.024750936776399612, + 0.023449258878827095, + 0.017914550378918648, + 0.0024798638187348843, + 0.003018894698470831, + 0.003742630360648036, + 0.004744092468172312, + 0.002067855792120099, + 0.004582489375025034, + 0.00513919023796916, + 0.004283709917217493, + 0.02415519952774048, + 0.018290627747774124, + 0.025575723499059677, + 0.0370325967669487, + 0.013682826422154903, + 0.017533641308546066, + 0.032943978905677795, + 0.033254921436309814, + 0.0006266467389650643, + 0.0007088617421686649, + 0.0010331925004720688, + 0.0011343725491315126, + 0.001038859598338604, + 0.0010254539083689451, + 0.0010463092476129532, + 0.0008646103087812662, + 0.009571852162480354, + 0.008181709796190262, + 0.008237548172473907, + 0.00873134471476078, + 0.006489500403404236, + 0.008677775040268898, + 0.006003898102790117, + 0.007842779159545898, + 0.005133641883730888, + 0.00813375785946846, + 0.004485469777137041, + 0.005258016753941774, + 0.0033030863851308823, + 0.006128992419689894, + 0.0027451615314930677, + 0.003762083826586604, + 0.0011837916681542993, + 0.0014358325861394405, + 0.002733472967520356, + 0.002136245835572481, + 0.002031927462667227, + 0.0015108337393030524, + 0.0018602994969114661, + 0.0024019198026508093, + 0.02602948620915413, + 0.015145908109843731, + 0.11617125570774078, + 0.04255270957946777, + 0.02251548320055008, + 0.02325553260743618, + 0.01792876236140728, + 0.10003317147493362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.014186837710440159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, steady shot, shot at eye level, captures a tranquil, sun-drenched scene. An elderly man, his face etched with soft wrinkles, his silver-white hair gleaming in the light, methodically mows a vast, vibrant green lawn. Wearing a faded light blue work shirt with sleeves rolled up to the elbows and simple brown trousers, he maintains a slightly bent posture as he steadily and deliberately pushes a classic, dark green manual reel mower. With each push, the mower's blades rotate with a quiet, mechanical whir, neatly mowing the emerald green grass and leaving a perfectly straight, slightly darkened path in their wake. In the background, a row of charming storybook cottages sit beneath a clear sky, their white stone walls adorned with climbing roses and their window sills overflowing with flower boxes. Their most striking feature is their thick, intricately tiered roofs, which glow a warm golden-brown in the afternoon sun. The warm light bathes the entire landscape, casting long, soft shadows and creating an idyllic atmosphere of peace and eternal, quiet contentment.", + "seed": 32261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_9/index_7300_seed_32261_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.20231310375966133, + "vlm_logprob": -0.20231310375966133, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.20231310375966133, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.7054263353347778 + ], + "opposite_logprob_tokens": [ + -1.0804263353347778 + ], + "logprob_reward": -0.7054263353347778, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.7054263353347778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41997215151786804, + 0.04761137440800667, + 0.20142903923988342, + 0.040730830281972885, + 0.03777644410729408, + 0.025560736656188965, + 0.04147053509950638, + 0.05506451055407524, + 0.1186579093337059, + 0.011726434342563152 + ], + "credit_map": [ + 0.015521089546382427, + 0.01183773297816515, + 0.051742974668741226, + 0.03629738837480545, + 0.020821405574679375, + 0.03192345052957535, + 0.1706705540418625, + 0.08115757256746292, + 0.002498800167813897, + 0.003389539662748575, + 0.0028359866701066494, + 0.0031127710826694965, + 0.003804251551628113, + 0.003306020051240921, + 0.015687020495533943, + 0.01297698449343443, + 0.02253216691315174, + 0.015837039798498154, + 0.02323618158698082, + 0.029404960572719574, + 0.010927863419055939, + 0.014488128945231438, + 0.03398475795984268, + 0.051017940044403076, + 0.006813178304582834, + 0.003508789697661996, + 0.004003563430160284, + 0.005764931906014681, + 0.00252213841304183, + 0.003537286538630724, + 0.007882407866418362, + 0.006698535289615393, + 0.003747645067051053, + 0.004208555445075035, + 0.004830582533031702, + 0.005084198899567127, + 0.0025568311102688313, + 0.004570423625409603, + 0.004014718346297741, + 0.008763488382101059, + 0.0027156537398695946, + 0.002830501878634095, + 0.0038709728978574276, + 0.0035120665561407804, + 0.002070213435217738, + 0.003463489469140768, + 0.0024528244975954294, + 0.004645014647394419, + 0.003254269016906619, + 0.0033915885724127293, + 0.004174497444182634, + 0.004639572463929653, + 0.003762820502743125, + 0.007183107081800699, + 0.0068501937203109264, + 0.008214488625526428, + 0.004462826531380415, + 0.007382438518106937, + 0.007623045705258846, + 0.007074948400259018, + 0.00507853738963604, + 0.00685976492241025, + 0.006831742357462645, + 0.009751209057867527, + 0.0046091871336102486, + 0.0038133126217871904, + 0.013884302228689194, + 0.01786763407289982, + 0.0053248475305736065, + 0.007672794163227081, + 0.04373692721128464, + 0.021748902276158333, + 0.0010627522133290768, + 0.000952508591581136, + 0.0013165954733267426, + 0.001459371647797525, + 0.0012013071682304144, + 0.001326792174950242, + 0.0022564835380762815, + 0.002150624291971326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.7054263353347778 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28132811188697815 + ], + "opposite_logprob_tokens": [ + -1.4063280820846558 + ], + "logprob_reward": -0.28132811188697815, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28132811188697815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2116575539112091, + 0.03754754737019539, + 0.11214574426412582, + 0.12113136053085327, + 0.07038938254117966, + 0.03212607279419899, + 0.04331481084227562, + 0.03777696192264557, + 0.311907023191452, + 0.022003518417477608 + ], + "credit_map": [ + 0.014902944676578045, + 0.010448971763253212, + 0.04068769887089729, + 0.02862389199435711, + 0.017193347215652466, + 0.024327531456947327, + 0.04971422627568245, + 0.025758950039744377, + 0.002520578447729349, + 0.0023996310774236917, + 0.003962923306971788, + 0.0045077004469931126, + 0.002884797053411603, + 0.003804767969995737, + 0.00746621610596776, + 0.010000934824347496, + 0.012478995136916637, + 0.011423751711845398, + 0.01264939270913601, + 0.01740308478474617, + 0.008092056959867477, + 0.010213601402938366, + 0.013899540528655052, + 0.025985321030020714, + 0.021832261234521866, + 0.011554586701095104, + 0.011953093111515045, + 0.015169313177466393, + 0.011416446417570114, + 0.01217183843255043, + 0.0204556193202734, + 0.016578203067183495, + 0.007148025091737509, + 0.006838575936853886, + 0.007449943572282791, + 0.007001303136348724, + 0.00507363211363554, + 0.012029142118990421, + 0.009002616629004478, + 0.015846140682697296, + 0.0024177490267902613, + 0.002171530853956938, + 0.006478969939053059, + 0.006708826869726181, + 0.0020318403840065002, + 0.0024804240092635155, + 0.0027713836170732975, + 0.007065350655466318, + 0.0030424222350120544, + 0.0027321751695126295, + 0.0060017770156264305, + 0.006258600857108831, + 0.0028940048068761826, + 0.004282095935195684, + 0.009136888198554516, + 0.008966845460236073, + 0.0033941403962671757, + 0.0039917584508657455, + 0.004496418405324221, + 0.005562785547226667, + 0.003063636366277933, + 0.003606516867876053, + 0.007802203763276339, + 0.005859502591192722, + 0.010104355402290821, + 0.008645658381283283, + 0.03413693979382515, + 0.04608934000134468, + 0.014909680932760239, + 0.022861648350954056, + 0.12143558263778687, + 0.05372384563088417, + 0.0016674596117809415, + 0.0015007054898887873, + 0.0017994156805798411, + 0.0026890968438237906, + 0.0022592374589294195, + 0.0020699065644294024, + 0.006161428987979889, + 0.003856266848742962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.28132811188697815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012555602006614208 + ], + "opposite_logprob_tokens": [ + -4.3875555992126465 + ], + "logprob_reward": -0.012555602006614208, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012555602006614208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3297286033630371, + 0.03861144185066223, + 0.14788483083248138, + 0.0964462161064148, + 0.042089104652404785, + 0.024269862100481987, + 0.046807389706373215, + 0.03144005313515663, + 0.22579969465732574, + 0.016922766342759132 + ], + "credit_map": [ + 0.017313726246356964, + 0.011102848686277866, + 0.039640337228775024, + 0.04801195487380028, + 0.01862536184489727, + 0.028339743614196777, + 0.07671884447336197, + 0.08997579663991928, + 0.00292714755050838, + 0.003101659007370472, + 0.004742852877825499, + 0.005333753768354654, + 0.0036862166598439217, + 0.004062708001583815, + 0.005985509138554335, + 0.008771591819822788, + 0.019762380048632622, + 0.016734149307012558, + 0.015972597524523735, + 0.02234133705496788, + 0.010220232419669628, + 0.012481290847063065, + 0.017393972724676132, + 0.032978881150484085, + 0.017558163031935692, + 0.008688405156135559, + 0.009453007020056248, + 0.012295018881559372, + 0.009323709644377232, + 0.010087654925882816, + 0.015714354813098907, + 0.013325902633368969, + 0.0029454694595187902, + 0.0042405142448842525, + 0.006509358994662762, + 0.006290408316999674, + 0.0033605729695409536, + 0.006475348491221666, + 0.005359187256544828, + 0.006908246781677008, + 0.003809492103755474, + 0.003711884608492255, + 0.0028667219448834658, + 0.002126058330759406, + 0.002074153395369649, + 0.002748047700151801, + 0.00268360273912549, + 0.004249901510775089, + 0.0032474491745233536, + 0.0031794982496649027, + 0.006318902131170034, + 0.0066367629915475845, + 0.0036208743695169687, + 0.006055545527487993, + 0.007615531329065561, + 0.010132822208106518, + 0.0030357730574905872, + 0.003255661344155669, + 0.004708378575742245, + 0.00500982441008091, + 0.002393196802586317, + 0.0032410346902906895, + 0.0054347217082977295, + 0.004361460451036692, + 0.007644420024007559, + 0.005872897803783417, + 0.025001224130392075, + 0.03334685042500496, + 0.01059892401099205, + 0.015362212434411049, + 0.08881231397390366, + 0.03916085511445999, + 0.0012490777298808098, + 0.0016492350259795785, + 0.0013784043258056045, + 0.002149851992726326, + 0.0018751267343759537, + 0.001980192493647337, + 0.0038756972644478083, + 0.002765180077403784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.012555602006614208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003620974952355027 + ], + "opposite_logprob_tokens": [ + -5.6286211013793945 + ], + "logprob_reward": -0.003620974952355027, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003620974952355027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3540096580982208, + 0.04926357790827751, + 0.10175412893295288, + 0.0839867889881134, + 0.04094453528523445, + 0.0475989393889904, + 0.03496437892317772, + 0.03968794271349907, + 0.23089855909347534, + 0.01689150184392929 + ], + "credit_map": [ + 0.015800390392541885, + 0.011003026738762856, + 0.029499176889657974, + 0.04682378098368645, + 0.015427612699568272, + 0.02164389006793499, + 0.11060845106840134, + 0.10320332646369934, + 0.002627713605761528, + 0.0028354362584650517, + 0.004518680274486542, + 0.005114906933158636, + 0.0033751355949789286, + 0.002955643692985177, + 0.012452446855604649, + 0.015383611433207989, + 0.012815380468964577, + 0.011144022457301617, + 0.01087096706032753, + 0.014865390956401825, + 0.007914233021438122, + 0.009312054142355919, + 0.011907546781003475, + 0.022924533113837242, + 0.0142098693177104, + 0.007163177244365215, + 0.008720590732991695, + 0.011267423629760742, + 0.00763358362019062, + 0.008665213361382484, + 0.014891168102622032, + 0.011435757391154766, + 0.0034680303651839495, + 0.003888506442308426, + 0.004271332174539566, + 0.004359516780823469, + 0.0030657369643449783, + 0.007238616701215506, + 0.005262254737317562, + 0.009390536695718765, + 0.007899563759565353, + 0.008020619861781597, + 0.005207097623497248, + 0.004802740179002285, + 0.004167985636740923, + 0.00602076668292284, + 0.00505876773968339, + 0.0064213951118290424, + 0.0031688581220805645, + 0.002702076220884919, + 0.005637588910758495, + 0.005088969133794308, + 0.0023941555991768837, + 0.0037163852248340845, + 0.0037730501499027014, + 0.008483292534947395, + 0.003732826793566346, + 0.003967520780861378, + 0.006404714658856392, + 0.006345395930111408, + 0.0031922082416713238, + 0.005046758335083723, + 0.005453263409435749, + 0.005545256659388542, + 0.007728613447397947, + 0.006147639825940132, + 0.02558092400431633, + 0.03424249589443207, + 0.01096614170819521, + 0.015770504251122475, + 0.09056335687637329, + 0.039898864924907684, + 0.0013768449425697327, + 0.00191312818787992, + 0.0014788458356633782, + 0.0021855859085917473, + 0.001729101873934269, + 0.002131676534190774, + 0.003530023153871298, + 0.0025462957564741373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.003620974952355027 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634494617581367 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634494617581367, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634494617581367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6494584679603577, + 0.035020723938941956, + 0.06053396314382553, + 0.023917729035019875, + 0.03091823309659958, + 0.020374519750475883, + 0.03764929249882698, + 0.028376512229442596, + 0.09005232900381088, + 0.023698272183537483 + ], + "credit_map": [ + 0.014217170886695385, + 0.012616670690476894, + 0.04446098208427429, + 0.03981701657176018, + 0.019957680255174637, + 0.022002562880516052, + 0.29805952310562134, + 0.19832678139209747, + 0.0020430032163858414, + 0.002581891370937228, + 0.003918694332242012, + 0.0046293181367218494, + 0.0021985587663948536, + 0.0031264962162822485, + 0.009572668001055717, + 0.006950095761567354, + 0.005630482919514179, + 0.0044966223649680614, + 0.004377409815788269, + 0.009593887254595757, + 0.005490667652338743, + 0.0053637828677892685, + 0.016092563048005104, + 0.009488546289503574, + 0.003020544769242406, + 0.001838877098634839, + 0.0028676381334662437, + 0.0036358716897666454, + 0.002300051972270012, + 0.002531968057155609, + 0.004390016198158264, + 0.0033327611163258553, + 0.0043029217049479485, + 0.003869903739541769, + 0.004144101403653622, + 0.0037065420765429735, + 0.0021684260573238134, + 0.0026711318641901016, + 0.004311809316277504, + 0.00574339646846056, + 0.002965368563309312, + 0.002779706148430705, + 0.0027365267742425203, + 0.0024593384005129337, + 0.0017740912735462189, + 0.002427005907520652, + 0.001937342225573957, + 0.003295141039416194, + 0.003584602614864707, + 0.0035111126489937305, + 0.004303210414946079, + 0.005438046995550394, + 0.002761675976216793, + 0.004313528537750244, + 0.008508932776749134, + 0.005228185560554266, + 0.0024194035213440657, + 0.00348311266861856, + 0.003062337404116988, + 0.003267609514296055, + 0.0024415310472249985, + 0.0038185250014066696, + 0.005129839759320021, + 0.004754153545945883, + 0.003702598623931408, + 0.0032184056472033262, + 0.009130763821303844, + 0.012930148281157017, + 0.004981475416570902, + 0.005481123924255371, + 0.035920239984989166, + 0.014687571674585342, + 0.002430474618449807, + 0.002545570954680443, + 0.00222576386295259, + 0.003927840851247311, + 0.0022456396836787462, + 0.0031744258012622595, + 0.002883381675928831, + 0.0042651742696762085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.008634494617581367 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, steady shot, shot at eye level, captures a tranquil, sun-drenched scene. An elderly man, his face etched with soft wrinkles, his silver-white hair gleaming in the light, methodically mows a vast, vibrant green lawn. Wearing a faded light blue work shirt with sleeves rolled up to the elbows and simple brown trousers, he maintains a slightly bent posture as he steadily and deliberately pushes a classic, dark green manual reel mower. With each push, the mower's blades rotate with a quiet, mechanical whir, neatly mowing the emerald green grass and leaving a perfectly straight, slightly darkened path in their wake. In the background, a row of charming storybook cottages sit beneath a clear sky, their white stone walls adorned with climbing roses and their window sills overflowing with flower boxes. Their most striking feature is their thick, intricately tiered roofs, which glow a warm golden-brown in the afternoon sun. The warm light bathes the entire landscape, casting long, soft shadows and creating an idyllic atmosphere of peace and eternal, quiet contentment.", + "seed": 32262, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_9/index_7300_seed_32262_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1449491621227935, + "vlm_logprob": -0.1449491621227935, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.1449491621227935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6337002515792847 + ], + "opposite_logprob_tokens": [ + -1.5087002515792847 + ], + "logprob_reward": -0.6337002515792847, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.6337002515792847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2485756129026413, + 0.04846179112792015, + 0.03705349564552307, + 0.09276082366704941, + 0.04783053323626518, + 0.03691250458359718, + 0.06034033000469208, + 0.14461329579353333, + 0.1740187108516693, + 0.10943295061588287 + ], + "credit_map": [ + 0.021277325227856636, + 0.017086509615182877, + 0.038947463035583496, + 0.0595121793448925, + 0.015902964398264885, + 0.016278743743896484, + 0.035990409553050995, + 0.04358001425862312, + 0.006466516759246588, + 0.0038165643345564604, + 0.007621111813932657, + 0.0070467316545546055, + 0.0045288545079529285, + 0.007720898371189833, + 0.0053605991415679455, + 0.005900515243411064, + 0.0055395751260221004, + 0.003908703103661537, + 0.004222449846565723, + 0.005549246445298195, + 0.0038786791265010834, + 0.004746390972286463, + 0.004685829393565655, + 0.004522622562944889, + 0.012562352232635021, + 0.006868371739983559, + 0.012631088495254517, + 0.015383119694888592, + 0.008938603103160858, + 0.01251065731048584, + 0.013113033026456833, + 0.010753597132861614, + 0.006349485367536545, + 0.004679256584495306, + 0.006246978882700205, + 0.007355355657637119, + 0.004486675839871168, + 0.005460930522531271, + 0.006486332509666681, + 0.006765521131455898, + 0.005708872806280851, + 0.0028429352678358555, + 0.005059757269918919, + 0.00547211617231369, + 0.004218786023557186, + 0.004594799131155014, + 0.004146616440266371, + 0.004868620075285435, + 0.004430333152413368, + 0.0036052961368113756, + 0.007236816454678774, + 0.014000984840095043, + 0.004160190001130104, + 0.003556713694706559, + 0.014056547544896603, + 0.009293446317315102, + 0.014404467307031155, + 0.009454378858208656, + 0.03690421208739281, + 0.030557163059711456, + 0.010515272617340088, + 0.00876980647444725, + 0.011575599201023579, + 0.022432386875152588, + 0.02527410350739956, + 0.01451812218874693, + 0.02089105360209942, + 0.024436818435788155, + 0.0251187551766634, + 0.013732739724218845, + 0.028276309370994568, + 0.02177080139517784, + 0.011941459961235523, + 0.006697733420878649, + 0.023480599746108055, + 0.016886264085769653, + 0.007181027438491583, + 0.007265263237059116, + 0.013791033066809177, + 0.022189565002918243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.6337002515792847 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003826647996902 + ], + "opposite_logprob_tokens": [ + -2.695038318634033 + ], + "logprob_reward": -0.07003826647996902, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003826647996902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.273456871509552, + 0.05440984666347504, + 0.04042298346757889, + 0.07002957910299301, + 0.03272011876106262, + 0.019808437675237656, + 0.04520463943481445, + 0.08942209184169769, + 0.24601365625858307, + 0.12851177155971527 + ], + "credit_map": [ + 0.016479982063174248, + 0.024708030745387077, + 0.04454823583364487, + 0.06431186944246292, + 0.020734112709760666, + 0.016467733308672905, + 0.040729641914367676, + 0.0454772524535656, + 0.005410146899521351, + 0.0043554166331887245, + 0.006304714363068342, + 0.01297654490917921, + 0.004145214334130287, + 0.004889743402600288, + 0.00766260689124465, + 0.008665459230542183, + 0.0034412883687764406, + 0.0030366904102265835, + 0.0044747162610292435, + 0.00946811493486166, + 0.003609882201999426, + 0.003443410387262702, + 0.006671539507806301, + 0.006277342326939106, + 0.011948310770094395, + 0.006860441993921995, + 0.0088686877861619, + 0.00835358165204525, + 0.005991293117403984, + 0.010286157950758934, + 0.008719298988580704, + 0.009001808241009712, + 0.004851778503507376, + 0.0023741363547742367, + 0.003634235356003046, + 0.0070937760174274445, + 0.002487951423972845, + 0.0038894768804311752, + 0.003978903405368328, + 0.004409859888255596, + 0.002673874609172344, + 0.0012178856413811445, + 0.0029576090164482594, + 0.0029491279274225235, + 0.001268526422791183, + 0.002015782753005624, + 0.003883825382217765, + 0.002841806970536709, + 0.0027902047149837017, + 0.002409783424809575, + 0.006268572527915239, + 0.012408064678311348, + 0.0027127061039209366, + 0.0026018768548965454, + 0.009687556885182858, + 0.006325872614979744, + 0.00909692794084549, + 0.00775401433929801, + 0.009533636271953583, + 0.009682850912213326, + 0.014170559123158455, + 0.009634717367589474, + 0.017722712829709053, + 0.011826667934656143, + 0.02278757654130459, + 0.01984846219420433, + 0.0419577993452549, + 0.03594213351607323, + 0.023141853511333466, + 0.02743549831211567, + 0.03358807414770126, + 0.04131224751472473, + 0.006742799188941717, + 0.005031687673181295, + 0.04086525738239288, + 0.010441816411912441, + 0.007473233621567488, + 0.006559307686984539, + 0.011588778346776962, + 0.03980889171361923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.07003826647996902 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751937326043844 + ], + "opposite_logprob_tokens": [ + -5.006752014160156 + ], + "logprob_reward": -0.006751937326043844, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751937326043844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1396016776561737, + 0.036227867007255554, + 0.03319007158279419, + 0.026887943968176842, + 0.02365228347480297, + 0.019339963793754578, + 0.015532740391790867, + 0.14359800517559052, + 0.3197677731513977, + 0.2422017902135849 + ], + "credit_map": [ + 0.009543880820274353, + 0.011987277306616306, + 0.018887905403971672, + 0.0339844673871994, + 0.010813440196216106, + 0.01019063126295805, + 0.02065267786383629, + 0.02354137971997261, + 0.0030292621813714504, + 0.0035278841387480497, + 0.00515513913705945, + 0.007601428311318159, + 0.0030581646133214235, + 0.002852382604032755, + 0.0055889650247991085, + 0.005414638202637434, + 0.0027334431651979685, + 0.0024254850577563047, + 0.003406903240829706, + 0.008818896487355232, + 0.0029507866129279137, + 0.0026376002933830023, + 0.00494288420304656, + 0.00527406670153141, + 0.003555163973942399, + 0.0022449966054409742, + 0.003836224088445306, + 0.004414556082338095, + 0.001959068700671196, + 0.0037045420613139868, + 0.004016007296741009, + 0.0031573823653161526, + 0.003697805106639862, + 0.0017780716298148036, + 0.002761788433417678, + 0.004035331774502993, + 0.002687149913981557, + 0.0025676314253360033, + 0.0025806413032114506, + 0.0035438616760075092, + 0.0030687968246638775, + 0.0012144335778430104, + 0.0022426787763834, + 0.0022218641825020313, + 0.002119687618687749, + 0.0020263097248971462, + 0.0037484217900782824, + 0.0026977688539773226, + 0.0015259389765560627, + 0.001202932558953762, + 0.0020836531184613705, + 0.0026274609845131636, + 0.0014614569954574108, + 0.0021299824584275484, + 0.0025474680587649345, + 0.0019538463093340397, + 0.012706362642347813, + 0.012939157895743847, + 0.021702338010072708, + 0.023397032171487808, + 0.01836930587887764, + 0.019273752346634865, + 0.01662192866206169, + 0.018588107079267502, + 0.027778981253504753, + 0.02878294326364994, + 0.053290095180273056, + 0.0499449297785759, + 0.03471241518855095, + 0.033553604036569595, + 0.04545816406607628, + 0.046246595680713654, + 0.01061598677188158, + 0.00960218720138073, + 0.08225847035646439, + 0.017911521717905998, + 0.016234496608376503, + 0.012533118017017841, + 0.0175674706697464, + 0.07547850906848907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.006751937326043844 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001718236249871552 + ], + "opposite_logprob_tokens": [ + -6.376718044281006 + ], + "logprob_reward": -0.001718236249871552, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001718236249871552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17166396975517273, + 0.0534808374941349, + 0.07235153764486313, + 0.06227535009384155, + 0.09619782865047455, + 0.050108473747968674, + 0.02236984111368656, + 0.062324993312358856, + 0.18031136691570282, + 0.2289157658815384 + ], + "credit_map": [ + 0.011499986983835697, + 0.01664026826620102, + 0.02446727268397808, + 0.04275137558579445, + 0.012720666825771332, + 0.010355894453823566, + 0.02558298595249653, + 0.027645505964756012, + 0.006153017282485962, + 0.004081159830093384, + 0.006279815919697285, + 0.012080501765012741, + 0.004114006645977497, + 0.006085896398872137, + 0.006190691143274307, + 0.008495750837028027, + 0.006940900348126888, + 0.0047745113261044025, + 0.00934862531721592, + 0.01893739216029644, + 0.00498184934258461, + 0.004226038698107004, + 0.010800231248140335, + 0.012341991998255253, + 0.01032076682895422, + 0.004553134553134441, + 0.006747670937329531, + 0.012062803842127323, + 0.004473013337701559, + 0.005916287656873465, + 0.009624071419239044, + 0.008577599190175533, + 0.012664475478231907, + 0.004842793568968773, + 0.010649342089891434, + 0.02184543013572693, + 0.007045697886496782, + 0.006574926432222128, + 0.015518547967076302, + 0.01705661416053772, + 0.005101448390632868, + 0.002705346792936325, + 0.006605197209864855, + 0.006886476185172796, + 0.004079552832990885, + 0.00413011759519577, + 0.012925933115184307, + 0.007674402091652155, + 0.0024174018763005733, + 0.0016941583016887307, + 0.0028796133119612932, + 0.003965305630117655, + 0.0022061613854020834, + 0.0022764871828258038, + 0.0037177661433815956, + 0.003212948562577367, + 0.008344165049493313, + 0.004461704753339291, + 0.012201773934066296, + 0.009816057980060577, + 0.004365576896816492, + 0.007400046568363905, + 0.005346932448446751, + 0.010388733819127083, + 0.025061683729290962, + 0.011701665818691254, + 0.02768256701529026, + 0.022929398342967033, + 0.013209919445216656, + 0.020628226920962334, + 0.026302585378289223, + 0.03279532119631767, + 0.012853814288973808, + 0.010689685121178627, + 0.06789473444223404, + 0.025335876271128654, + 0.012674189172685146, + 0.013229682110249996, + 0.02101868949830532, + 0.06521911174058914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.001718236249871552 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253711897879839 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253711897879839, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253711897879839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2724432349205017, + 0.08845009654760361, + 0.05465841293334961, + 0.0410478375852108, + 0.0464591458439827, + 0.01806783489882946, + 0.023536913096904755, + 0.0451502688229084, + 0.10790988802909851, + 0.3022763729095459 + ], + "credit_map": [ + 0.03130532428622246, + 0.017849911004304886, + 0.03376752883195877, + 0.04337037354707718, + 0.033619772642850876, + 0.030214210972189903, + 0.04465034604072571, + 0.037665754556655884, + 0.011300859972834587, + 0.009270216338336468, + 0.00794946774840355, + 0.01967921480536461, + 0.007418009452521801, + 0.012494644150137901, + 0.008373774588108063, + 0.011963900178670883, + 0.0069584352895617485, + 0.006349506787955761, + 0.0055466326884925365, + 0.010245759040117264, + 0.005177551880478859, + 0.006922678556293249, + 0.006367047317326069, + 0.0070907981134951115, + 0.004559611435979605, + 0.004336790181696415, + 0.004674108698964119, + 0.006070081610232592, + 0.004837764427065849, + 0.006193271838128567, + 0.005571506451815367, + 0.004804701078683138, + 0.004586199764162302, + 0.00439990870654583, + 0.004546438809484243, + 0.0081204604357481, + 0.0057061705738306046, + 0.006636360194534063, + 0.006927223410457373, + 0.005536382552236319, + 0.0026501549873501062, + 0.0020047526340931654, + 0.002557932399213314, + 0.002012144075706601, + 0.0018903929740190506, + 0.002144637517631054, + 0.0026107225567102432, + 0.0021970970556139946, + 0.0022487365640699863, + 0.0011793813901022077, + 0.00303499773144722, + 0.005161983426660299, + 0.0014947073068469763, + 0.0019030877156183124, + 0.005088392179459333, + 0.0034256235230714083, + 0.005045195575803518, + 0.0033549852669239044, + 0.007791890297085047, + 0.005790902301669121, + 0.004350490868091583, + 0.004118382930755615, + 0.008231221698224545, + 0.006467198021709919, + 0.016539765521883965, + 0.007824867032468319, + 0.011807716451585293, + 0.011935809627175331, + 0.012235721573233604, + 0.010307472199201584, + 0.019739529117941856, + 0.017519008368253708, + 0.021449051797389984, + 0.015618914738297462, + 0.08108291029930115, + 0.034615300595760345, + 0.0179597195237875, + 0.020311987027525902, + 0.030566543340682983, + 0.08067190647125244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.01253711897879839 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, steady shot, shot at eye level, captures a tranquil, sun-drenched scene. An elderly man, his face etched with soft wrinkles, his silver-white hair gleaming in the light, methodically mows a vast, vibrant green lawn. Wearing a faded light blue work shirt with sleeves rolled up to the elbows and simple brown trousers, he maintains a slightly bent posture as he steadily and deliberately pushes a classic, dark green manual reel mower. With each push, the mower's blades rotate with a quiet, mechanical whir, neatly mowing the emerald green grass and leaving a perfectly straight, slightly darkened path in their wake. In the background, a row of charming storybook cottages sit beneath a clear sky, their white stone walls adorned with climbing roses and their window sills overflowing with flower boxes. Their most striking feature is their thick, intricately tiered roofs, which glow a warm golden-brown in the afternoon sun. The warm light bathes the entire landscape, casting long, soft shadows and creating an idyllic atmosphere of peace and eternal, quiet contentment.", + "seed": 32263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000126/group_9/index_7300_seed_32263_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14559986935928465, + "vlm_logprob": -0.14559986935928465, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.14559986935928465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5219885110855103 + ], + "opposite_logprob_tokens": [ + -1.5219885110855103 + ], + "logprob_reward": -0.5219885110855103, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.5219885110855103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14310938119888306, + 0.04592619836330414, + 0.027770332992076874, + 0.043385062366724014, + 0.04155680537223816, + 0.06891214102506638, + 0.11806558072566986, + 0.03844592347741127, + 0.4517931342124939, + 0.021035395562648773 + ], + "credit_map": [ + 0.012757883407175541, + 0.011038575321435928, + 0.028720760717988014, + 0.025072019547224045, + 0.012341877445578575, + 0.014401504769921303, + 0.014450892806053162, + 0.024325864389538765, + 0.003134398953989148, + 0.0024046883918344975, + 0.006338660605251789, + 0.016389280557632446, + 0.0021509972866624594, + 0.00189547473564744, + 0.006169688887894154, + 0.007443005684763193, + 0.0018196017481386662, + 0.001893451320938766, + 0.0034846183843910694, + 0.005873204208910465, + 0.004060190171003342, + 0.0023711964022368193, + 0.00476008839905262, + 0.003507984336465597, + 0.00525796040892601, + 0.002644780557602644, + 0.008889978751540184, + 0.005937878508120775, + 0.0026351541746407747, + 0.0045180250890553, + 0.004822343587875366, + 0.008678944781422615, + 0.005685725249350071, + 0.003495879704132676, + 0.010647940449416637, + 0.00471567502245307, + 0.001939824316650629, + 0.0036025592125952244, + 0.00447767972946167, + 0.006991524249315262, + 0.005490648094564676, + 0.0048574586398899555, + 0.009549125097692013, + 0.016648320481181145, + 0.007719482760876417, + 0.0034647604916244745, + 0.013496018946170807, + 0.007686326280236244, + 0.01258088555186987, + 0.01161455363035202, + 0.020933035761117935, + 0.020544469356536865, + 0.010495097376406193, + 0.008340817876160145, + 0.012748859822750092, + 0.020807864144444466, + 0.0039744507521390915, + 0.0030653513967990875, + 0.00809968076646328, + 0.006622541230171919, + 0.002920896979048848, + 0.002734061097726226, + 0.0038733321707695723, + 0.0071556102484464645, + 0.015380061231553555, + 0.020585549995303154, + 0.02717575617134571, + 0.15558937191963196, + 0.019241224974393845, + 0.01910904422402382, + 0.025051040574908257, + 0.1696610450744629, + 0.0010250492487102747, + 0.0012089787051081657, + 0.0017902402905747294, + 0.0069032530300319195, + 0.0013328728964552283, + 0.001219552243128419, + 0.0015263200039044023, + 0.0060291290283203125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.5219885110855103 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989513278007507 + ], + "opposite_logprob_tokens": [ + -1.8048951625823975 + ], + "logprob_reward": -0.17989513278007507, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989513278007507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5061421394348145, + 0.02744227461516857, + 0.03665810078382492, + 0.011063341982662678, + 0.023166904225945473, + 0.06567247211933136, + 0.03430207818746567, + 0.015780556946992874, + 0.26491478085517883, + 0.014857378788292408 + ], + "credit_map": [ + 0.0436539351940155, + 0.026377560570836067, + 0.10727153718471527, + 0.07577475905418396, + 0.04197671636939049, + 0.040405381470918655, + 0.0752733126282692, + 0.09540894627571106, + 0.0029865747783333063, + 0.0024887588806450367, + 0.003947929944843054, + 0.006958008278161287, + 0.002255526836961508, + 0.0022270320914685726, + 0.0033865489531308413, + 0.003191892523318529, + 0.004127200227230787, + 0.0030042470898479223, + 0.00795233715325594, + 0.004859724082052708, + 0.002974909031763673, + 0.0035030997823923826, + 0.0037513552233576775, + 0.006485228426754475, + 0.0011202616151422262, + 0.0013408169616013765, + 0.0013814994599670172, + 0.0016337524866685271, + 0.0013224786380305886, + 0.0006799704278819263, + 0.0024834461510181427, + 0.0011011166498064995, + 0.002354755299165845, + 0.00228899042122066, + 0.0024192214477807283, + 0.003407006384804845, + 0.0022105558309704065, + 0.0019496289314702153, + 0.0057080877013504505, + 0.002828657627105713, + 0.00466179521754384, + 0.0042604622431099415, + 0.008051291108131409, + 0.01310744322836399, + 0.008816526271402836, + 0.0038190516643226147, + 0.015685226768255234, + 0.007270670961588621, + 0.004174590576440096, + 0.0026035134214907885, + 0.006384611129760742, + 0.005054962355643511, + 0.0034943127539008856, + 0.0028343675658106804, + 0.00409936485812068, + 0.00565635459497571, + 0.0011440811213105917, + 0.0010871379636228085, + 0.002087267115712166, + 0.0017011385643854737, + 0.001779212267138064, + 0.0009569570538587868, + 0.004496355541050434, + 0.002528408309444785, + 0.01182218175381422, + 0.02635100483894348, + 0.022715220227837563, + 0.0783616155385971, + 0.03323573246598244, + 0.009834444150328636, + 0.030728798359632492, + 0.051865797489881516, + 0.0009676828631199896, + 0.0008437273208983243, + 0.0015427499311044812, + 0.003446087474003434, + 0.0020922617986798286, + 0.0008564009913243353, + 0.0024487487971782684, + 0.002659720368683338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.17989513278007507 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255148183554411 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.01255148183554411, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255148183554411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39721810817718506, + 0.03312274068593979, + 0.0384192019701004, + 0.019249869510531425, + 0.06780178844928741, + 0.08865124732255936, + 0.0689009502530098, + 0.021160917356610298, + 0.22672279179096222, + 0.03875245526432991 + ], + "credit_map": [ + 0.03327205777168274, + 0.019539382308721542, + 0.08499477803707123, + 0.05917929857969284, + 0.0334787592291832, + 0.02973545528948307, + 0.06604188680648804, + 0.07097642868757248, + 0.0022108168341219425, + 0.0020106895826756954, + 0.004399070516228676, + 0.011017435230314732, + 0.0023364825174212456, + 0.002358329249545932, + 0.00408929493278265, + 0.004700621590018272, + 0.003336879890412092, + 0.0024424672592431307, + 0.0072664315812289715, + 0.007652141619473696, + 0.002461221069097519, + 0.004085191059857607, + 0.004492064472287893, + 0.0066828010603785515, + 0.0018956981366500258, + 0.0014757838798686862, + 0.0030129856895655394, + 0.002219083020463586, + 0.002127909567207098, + 0.001741215237416327, + 0.004218947608023882, + 0.0025582436937838793, + 0.005918582901358604, + 0.004496206063777208, + 0.012640934437513351, + 0.00969205517321825, + 0.007050263229757547, + 0.004738970194011927, + 0.013734875246882439, + 0.00952989887446165, + 0.006380471400916576, + 0.005240556783974171, + 0.011071635410189629, + 0.01648237369954586, + 0.011936998926103115, + 0.004878553561866283, + 0.022251782938838005, + 0.010408862493932247, + 0.006634120363742113, + 0.004090936854481697, + 0.01313097309321165, + 0.012578236870467663, + 0.007448660675436258, + 0.005810064729303122, + 0.008098676800727844, + 0.011109274812042713, + 0.0022910400293767452, + 0.0016888787504285574, + 0.0035209404304623604, + 0.004546890966594219, + 0.0013826903887093067, + 0.0016939629567787051, + 0.002065663691610098, + 0.003970845602452755, + 0.013549989089369774, + 0.016054773703217506, + 0.0197259783744812, + 0.03757421299815178, + 0.03140377998352051, + 0.013785652816295624, + 0.034508027136325836, + 0.0601203553378582, + 0.001542462152428925, + 0.0019140089862048626, + 0.0035897886846214533, + 0.014140443876385689, + 0.0019116102484986186, + 0.0024539006408303976, + 0.0047049992717802525, + 0.008495237678289413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.01255148183554411 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002493369858711958 + ], + "opposite_logprob_tokens": [ + -6.002493381500244 + ], + "logprob_reward": -0.002493369858711958, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002493369858711958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13144555687904358, + 0.04466085880994797, + 0.04096971079707146, + 0.03883789852261543, + 0.05072842165827751, + 0.024767311289906502, + 0.06046553701162338, + 0.030597779899835587, + 0.5440512299537659, + 0.033475589007139206 + ], + "credit_map": [ + 0.01203388161957264, + 0.011013713665306568, + 0.021639831364154816, + 0.018258167430758476, + 0.014094850048422813, + 0.011754695326089859, + 0.022507235407829285, + 0.020143186673521996, + 0.003951495047658682, + 0.0032676381524652243, + 0.007891083136200905, + 0.009457887150347233, + 0.0031789077911525965, + 0.002926806453615427, + 0.007398300338536501, + 0.006588739342987537, + 0.0035123208072036505, + 0.002931231865659356, + 0.0066822730004787445, + 0.005075456574559212, + 0.006298627704381943, + 0.0030752960592508316, + 0.006508549675345421, + 0.006885954178869724, + 0.005279554519802332, + 0.0018684726674109697, + 0.010370406322181225, + 0.0042609223164618015, + 0.001425191992893815, + 0.004025178961455822, + 0.0032207851763814688, + 0.008387390524148941, + 0.006536387372761965, + 0.0030697223264724016, + 0.011718666180968285, + 0.005624399520456791, + 0.00235943216830492, + 0.0047415466979146, + 0.006537663284689188, + 0.010140609927475452, + 0.0022868781816214323, + 0.0013016321463510394, + 0.003986310213804245, + 0.003912540152668953, + 0.003323313547298312, + 0.001269168104045093, + 0.005704556126147509, + 0.0029829139821231365, + 0.0064303153194487095, + 0.004873375408351421, + 0.008998215198516846, + 0.015763696283102036, + 0.0061475299298763275, + 0.0033651990815997124, + 0.007735452614724636, + 0.0071517545729875565, + 0.00208577373996377, + 0.0031907770317047834, + 0.004559728782624006, + 0.004236853681504726, + 0.003307214705273509, + 0.0021878285333514214, + 0.005778361577540636, + 0.005251245107501745, + 0.01924917846918106, + 0.02231185883283615, + 0.03531565144658089, + 0.1835235208272934, + 0.020514676347374916, + 0.02441871352493763, + 0.03669662028551102, + 0.202021062374115, + 0.0017063127597793937, + 0.002061172155663371, + 0.003350624116137624, + 0.01077236793935299, + 0.0018852361245080829, + 0.0022712235804647207, + 0.0040651424787938595, + 0.007363514509052038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.002493369858711958 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly man with silver-white hair and soft wrinkles visible mowing the lawn?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man wearing a faded light blue work shirt with sleeves rolled up and brown trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man pushing a classic dark green manual reel mower that leaves a straight, darkened path in the grass?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are storybook cottages with white stone walls, climbing roses, and thick golden-brown tiered roofs visible in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011070851236581802 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.011070851236581802, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011070851236581802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2202320694923401, + 0.06315609067678452, + 0.07911956310272217, + 0.0649527907371521, + 0.08335709571838379, + 0.07723643630743027, + 0.09765420854091644, + 0.06002328172326088, + 0.2140006422996521, + 0.040267784148454666 + ], + "credit_map": [ + 0.020906487479805946, + 0.022315211594104767, + 0.04184143617749214, + 0.023956499993801117, + 0.01844942942261696, + 0.02859068661928177, + 0.02134961262345314, + 0.04282271862030029, + 0.007013524882495403, + 0.007659238297492266, + 0.006271033082157373, + 0.007058616261929274, + 0.007262914441525936, + 0.0076241022907197475, + 0.012652590870857239, + 0.007614079397171736, + 0.0065366970375180244, + 0.007127966731786728, + 0.012556210160255432, + 0.0104572968557477, + 0.00954588782042265, + 0.00941652711480856, + 0.01275293156504631, + 0.010726052336394787, + 0.007652571890503168, + 0.008905733935534954, + 0.007597559131681919, + 0.014906412921845913, + 0.005088951904326677, + 0.00719861127436161, + 0.00597217446193099, + 0.007630790118128061, + 0.012085462920367718, + 0.015306158922612667, + 0.01430295780301094, + 0.010379280894994736, + 0.005343188066035509, + 0.007291544694453478, + 0.010421697050333023, + 0.008226812817156315, + 0.00784830842167139, + 0.007952971383929253, + 0.01112372800707817, + 0.015675591304898262, + 0.007729959208518267, + 0.004735851194709539, + 0.014063806273043156, + 0.008106229826807976, + 0.014246790669858456, + 0.012203291989862919, + 0.018757697194814682, + 0.015778852626681328, + 0.00735044339671731, + 0.00858708843588829, + 0.010123520158231258, + 0.010606531985104084, + 0.005560911260545254, + 0.007678487803786993, + 0.008826005272567272, + 0.012491297908127308, + 0.00422255415469408, + 0.006451602093875408, + 0.006245236843824387, + 0.008547193370759487, + 0.006977858021855354, + 0.012578828260302544, + 0.010888411663472652, + 0.0784960389137268, + 0.007595900911837816, + 0.009466948918998241, + 0.013774319551885128, + 0.07422234117984772, + 0.003954363986849785, + 0.005625386722385883, + 0.004943646024912596, + 0.0057762619107961655, + 0.003810481633991003, + 0.0038630093913525343, + 0.006941008381545544, + 0.005353630054742098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.011070851236581802 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_0/index_tensor([1029, 1029])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_0/index_tensor([1029, 1029])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aa0bb4a252330afc743b99247455d65078fbfa85 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_0/index_tensor([1029, 1029])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, held at eye level, tracks smoothly with the cyclist, then comes to an elegant, static stop as the cyclist comes to a stop. The cyclist, wearing a vibrant red lightweight windbreaker and a sophisticated matte black helmet, rides down a quiet, tree-lined asphalt street. The scene is bathed in the warm, golden light of late afternoon, casting long, soft shadows across the sidewalk. As the rider decides to stop, their pedaling deliberately slows, their leg movements becoming less frequent. The bicycle's wheel gradually slows, the spokes becoming more pronounced with each rotation, visually translating the gentle scraping of the tires on the road. The rider's initially resolute expression, focused on the road ahead, softens noticeably to a calm, reflective one. They bring the bike to a perfectly stable stop, release one foot, and keep it firmly planted on the ground for balance. In this final moment of silence, the tranquil setting, with the golden leaves of the overhanging trees swaying gently in the breeze, creates a dappled light effect that powerfully emphasizes the calm and contemplative atmosphere of the pause.", + "seed": 33428, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_0/index_1029_seed_33428_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.044160372391343114, + "vlm_logprob": -0.044160372391343114, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.044160372391343114, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17208132147789001 + ], + "opposite_logprob_tokens": [ + -2.297081232070923 + ], + "logprob_reward": -0.17208132147789001, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.17208132147789001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09191630780696869, + 0.04027552157640457, + 0.0867127776145935, + 0.11162477731704712, + 0.07020953297615051, + 0.2426472306251526, + 0.03100861795246601, + 0.022531209513545036, + 0.27554547786712646, + 0.02752857655286789 + ], + "credit_map": [ + 0.009828988462686539, + 0.012540911324322224, + 0.009951792657375336, + 0.00794246420264244, + 0.016124209389090538, + 0.01143165584653616, + 0.011623959988355637, + 0.012472324073314667, + 0.0039363703690469265, + 0.004492686130106449, + 0.00404663011431694, + 0.004890302661806345, + 0.005548466462641954, + 0.00624486431479454, + 0.004648648202419281, + 0.00646755238994956, + 0.007503244560211897, + 0.010022872127592564, + 0.00822026189416647, + 0.007820270024240017, + 0.025211984291672707, + 0.0074872225522994995, + 0.013232937082648277, + 0.007213988341391087, + 0.013893879018723965, + 0.020552434027194977, + 0.011482412926852703, + 0.010808623395860195, + 0.02273244969546795, + 0.009191029705107212, + 0.013905404135584831, + 0.009058546274900436, + 0.012584630399942398, + 0.007185013964772224, + 0.008327195420861244, + 0.00606759125366807, + 0.008344024419784546, + 0.01298340130597353, + 0.006522129289805889, + 0.008195546455681324, + 0.025232626125216484, + 0.04051349684596062, + 0.03403579071164131, + 0.019032413139939308, + 0.031717706471681595, + 0.027904001995921135, + 0.01848434843122959, + 0.045726846903562546, + 0.0028937675524502993, + 0.004525981843471527, + 0.0033517039846628904, + 0.004025483503937721, + 0.004232456907629967, + 0.0037574053276330233, + 0.003984494600445032, + 0.004237326793372631, + 0.0029457302298396826, + 0.0032486936543136835, + 0.0016975540202111006, + 0.003170018084347248, + 0.0033935212995857, + 0.0029201561119407415, + 0.002770061371847987, + 0.002385473810136318, + 0.030140923336148262, + 0.023214614018797874, + 0.027806103229522705, + 0.04966391623020172, + 0.028648637235164642, + 0.025856107473373413, + 0.05468863993883133, + 0.03552651405334473, + 0.0032393955625593662, + 0.003870121669024229, + 0.004095961805433035, + 0.00416873674839735, + 0.003811740316450596, + 0.0022629641462117434, + 0.003883525962010026, + 0.0021961298771202564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.17208132147789001 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009814697317779064 + ], + "opposite_logprob_tokens": [ + -4.634814739227295 + ], + "logprob_reward": -0.009814697317779064, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009814697317779064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07447992265224457, + 0.14088848233222961, + 0.02897508442401886, + 0.053953588008880615, + 0.07385730743408203, + 0.08686042577028275, + 0.06984540820121765, + 0.04420856758952141, + 0.35117751359939575, + 0.07575369626283646 + ], + "credit_map": [ + 0.007835633121430874, + 0.009422528557479382, + 0.009866274893283844, + 0.007152047473937273, + 0.012708290480077267, + 0.0097530921921134, + 0.011249108240008354, + 0.006492936983704567, + 0.00925017986446619, + 0.01458780001848936, + 0.014456321485340595, + 0.012673543766140938, + 0.037378840148448944, + 0.01623593270778656, + 0.01769041270017624, + 0.018615437671542168, + 0.0031061528716236353, + 0.0034471903927624226, + 0.003063819371163845, + 0.0040793223306536674, + 0.004521827679127455, + 0.0037775137461721897, + 0.0035467559937387705, + 0.0034324994776397943, + 0.008438092656433582, + 0.01590767316520214, + 0.0043449425138533115, + 0.0047304667532444, + 0.008197239600121975, + 0.0038010620046406984, + 0.005004480946809053, + 0.003529624780640006, + 0.007146292831748724, + 0.01349340844899416, + 0.007611510809510946, + 0.012214791029691696, + 0.01372623909264803, + 0.0067817880772054195, + 0.00672318646684289, + 0.006160085555166006, + 0.011858410201966763, + 0.01284143328666687, + 0.012590326368808746, + 0.007890621200203896, + 0.009091048501431942, + 0.01005901675671339, + 0.005981920287013054, + 0.01654764637351036, + 0.009687179699540138, + 0.007824225351214409, + 0.007354218978434801, + 0.012420827522873878, + 0.006524299271404743, + 0.009620334953069687, + 0.006297450512647629, + 0.010116868652403355, + 0.005096502136439085, + 0.008216213434934616, + 0.0031066229566931725, + 0.006386668886989355, + 0.005868770647794008, + 0.006664114072918892, + 0.004908144939690828, + 0.003961525857448578, + 0.031502675265073776, + 0.029978914186358452, + 0.0397355854511261, + 0.06688753515481949, + 0.045669812709093094, + 0.030506819486618042, + 0.07022209465503693, + 0.03667407110333443, + 0.007977958768606186, + 0.0064274100586771965, + 0.009558004327118397, + 0.011706158518791199, + 0.011848331429064274, + 0.0077595897018909454, + 0.013139134272933006, + 0.007337100803852081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.009814697317779064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864111352711916 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.00864111352711916, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864111352711916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3088412582874298, + 0.1749238520860672, + 0.07839224487543106, + 0.04459081217646599, + 0.03151344135403633, + 0.06738807260990143, + 0.03097454458475113, + 0.05084387958049774, + 0.14866268634796143, + 0.06386923789978027 + ], + "credit_map": [ + 0.039316605776548386, + 0.035483673214912415, + 0.05442507937550545, + 0.022407907992601395, + 0.03282921388745308, + 0.055245574563741684, + 0.03017517179250717, + 0.03895803168416023, + 0.01627277210354805, + 0.025676878169178963, + 0.015651719644665718, + 0.022896071895956993, + 0.016098354011774063, + 0.027022363618016243, + 0.02019348554313183, + 0.031112222000956535, + 0.016317376866936684, + 0.012831555679440498, + 0.008240984752774239, + 0.0068371957167983055, + 0.010121439583599567, + 0.010448768734931946, + 0.007713439874351025, + 0.005881479941308498, + 0.007996336556971073, + 0.007390295621007681, + 0.0062238313257694244, + 0.004008498042821884, + 0.00520192738622427, + 0.0050077359192073345, + 0.0033112906385213137, + 0.005450897384434938, + 0.004546055570244789, + 0.0035455499310046434, + 0.0035943796392530203, + 0.003868039231747389, + 0.004221629351377487, + 0.004980347119271755, + 0.002824061084538698, + 0.0039333803579211235, + 0.009960569441318512, + 0.010356117971241474, + 0.008396918885409832, + 0.007044163532555103, + 0.008688978850841522, + 0.009482398629188538, + 0.006119392346590757, + 0.007339526899158955, + 0.004103753715753555, + 0.003971798345446587, + 0.0033977192360907793, + 0.00497127091512084, + 0.002958110999315977, + 0.0037467340007424355, + 0.002733889501541853, + 0.005091266706585884, + 0.007530989125370979, + 0.004746215417981148, + 0.0042097861878573895, + 0.00516846077516675, + 0.005973482970148325, + 0.012917479500174522, + 0.005875039380043745, + 0.004422429483383894, + 0.014678380452096462, + 0.012969400733709335, + 0.01872483640909195, + 0.027228020131587982, + 0.019789595156908035, + 0.01568939909338951, + 0.0266437865793705, + 0.012939266860485077, + 0.010765650309622288, + 0.0051259491592645645, + 0.009402074851095676, + 0.009432689286768436, + 0.00865132361650467, + 0.006355409976094961, + 0.008099942468106747, + 0.006036196369677782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.00864111352711916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197885990142822 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197885990142822, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197885990142822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10282323509454727, + 0.14357993006706238, + 0.08816590905189514, + 0.06746013462543488, + 0.04001612961292267, + 0.04964236542582512, + 0.034219443798065186, + 0.02087819017469883, + 0.4317135810852051, + 0.02150108851492405 + ], + "credit_map": [ + 0.009078524075448513, + 0.00958132091909647, + 0.01581481844186783, + 0.008294188417494297, + 0.02118484489619732, + 0.01680142991244793, + 0.013540741056203842, + 0.008527377620339394, + 0.013853955082595348, + 0.019197527319192886, + 0.012124845758080482, + 0.017213603481650352, + 0.015770642086863518, + 0.021689636632800102, + 0.01639261282980442, + 0.027337126433849335, + 0.007078062742948532, + 0.01094452291727066, + 0.007332995068281889, + 0.009022783488035202, + 0.02567077986896038, + 0.006754984147846699, + 0.013723663985729218, + 0.007638127077370882, + 0.007916998118162155, + 0.01804582215845585, + 0.005319415591657162, + 0.008044480346143246, + 0.010947699658572674, + 0.004928607027977705, + 0.007971939630806446, + 0.0042851753532886505, + 0.004437518306076527, + 0.005413559731096029, + 0.0035136411897838116, + 0.004176226910203695, + 0.008165507577359676, + 0.004879269748926163, + 0.004974633455276489, + 0.004455776885151863, + 0.004500837996602058, + 0.007975440472364426, + 0.005994114093482494, + 0.006190190091729164, + 0.00612286664545536, + 0.00887439213693142, + 0.003965407609939575, + 0.006019122898578644, + 0.003922630567103624, + 0.004326402675360441, + 0.0038915001787245274, + 0.005144220311194658, + 0.0035914811305701733, + 0.004388125613331795, + 0.0030534928664565086, + 0.005901593714952469, + 0.0022247210144996643, + 0.0029328197706490755, + 0.0016408496303483844, + 0.002884404733777046, + 0.0026527345180511475, + 0.0034925469662994146, + 0.0029250702355057, + 0.0021250464487820864, + 0.043476253747940063, + 0.03753049671649933, + 0.04362058639526367, + 0.08039618283510208, + 0.05947331711649895, + 0.04524248465895653, + 0.07973620295524597, + 0.04223807528614998, + 0.002429007785394788, + 0.0021941470913589, + 0.004350592847913504, + 0.0026973478961735964, + 0.0025580248329788446, + 0.002072000876069069, + 0.002527313306927681, + 0.002672656672075391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.014197885990142822 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066843643784523 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016066843643784523, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066843643784523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17267046868801117, + 0.04428175836801529, + 0.04672524333000183, + 0.03609076514840126, + 0.06293164193630219, + 0.04711109772324562, + 0.10266588628292084, + 0.042212557047605515, + 0.35442158579826355, + 0.09088901430368423 + ], + "credit_map": [ + 0.01734214834868908, + 0.01832733489573002, + 0.020024241879582405, + 0.017439231276512146, + 0.03147270157933235, + 0.021874070167541504, + 0.024410275742411613, + 0.021780455484986305, + 0.0045636375434696674, + 0.0060812197625637054, + 0.004148085135966539, + 0.004493839107453823, + 0.0047060903161764145, + 0.007196683436632156, + 0.005458662286400795, + 0.007633542641997337, + 0.002811602782458067, + 0.005578786134719849, + 0.0037063739728182554, + 0.004339965991675854, + 0.011896423064172268, + 0.004089228343218565, + 0.007050135638564825, + 0.007252727635204792, + 0.0034483408089727163, + 0.006195457652211189, + 0.0030970568768680096, + 0.005348869599401951, + 0.005453805439174175, + 0.003038829891011119, + 0.005495388060808182, + 0.0040130168199539185, + 0.0077592465095222, + 0.008379354141652584, + 0.007205948233604431, + 0.006833995226770639, + 0.008944629691541195, + 0.009395329281687737, + 0.00793315377086401, + 0.006479989737272263, + 0.006176677066832781, + 0.006127560511231422, + 0.006494255270808935, + 0.005185085814446211, + 0.005405385512858629, + 0.005955302622169256, + 0.004135460592806339, + 0.007631372194737196, + 0.013509394600987434, + 0.011364010162651539, + 0.010893704369664192, + 0.015624873340129852, + 0.009554359130561352, + 0.01527372095733881, + 0.011517898179590702, + 0.014927918091416359, + 0.005744817666709423, + 0.00654098205268383, + 0.0031397680286318064, + 0.005953418556600809, + 0.00583814550191164, + 0.005735698156058788, + 0.004463274497538805, + 0.004796454217284918, + 0.05775851383805275, + 0.032409995794296265, + 0.03327742964029312, + 0.053844429552555084, + 0.051796771585941315, + 0.062162190675735474, + 0.03542119637131691, + 0.027751052752137184, + 0.018282970413565636, + 0.005589236505329609, + 0.012130816467106342, + 0.012931636534631252, + 0.007725735194981098, + 0.008833622559905052, + 0.010878748260438442, + 0.014516253024339676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.016066843643784523 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, held at eye level, tracks smoothly with the cyclist, then comes to an elegant, static stop as the cyclist comes to a stop. The cyclist, wearing a vibrant red lightweight windbreaker and a sophisticated matte black helmet, rides down a quiet, tree-lined asphalt street. The scene is bathed in the warm, golden light of late afternoon, casting long, soft shadows across the sidewalk. As the rider decides to stop, their pedaling deliberately slows, their leg movements becoming less frequent. The bicycle's wheel gradually slows, the spokes becoming more pronounced with each rotation, visually translating the gentle scraping of the tires on the road. The rider's initially resolute expression, focused on the road ahead, softens noticeably to a calm, reflective one. They bring the bike to a perfectly stable stop, release one foot, and keep it firmly planted on the ground for balance. In this final moment of silence, the tranquil setting, with the golden leaves of the overhanging trees swaying gently in the breeze, creates a dappled light effect that powerfully emphasizes the calm and contemplative atmosphere of the pause.", + "seed": 33429, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_0/index_1029_seed_33429_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04051000997424126, + "vlm_logprob": -0.04051000997424126, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.04051000997424126, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16739371418952942 + ], + "opposite_logprob_tokens": [ + -2.417393684387207 + ], + "logprob_reward": -0.16739371418952942, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.16739371418952942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4515652358531952, + 0.08489339053630829, + 0.03724345192313194, + 0.08748070150613785, + 0.050730906426906586, + 0.03908415883779526, + 0.09865672886371613, + 0.07963390648365021, + 0.053022149950265884, + 0.017689445987343788 + ], + "credit_map": [ + 0.043661393225193024, + 0.08666543662548065, + 0.04297582805156708, + 0.05085573345422745, + 0.10058070719242096, + 0.03996584191918373, + 0.05280635505914688, + 0.03405391052365303, + 0.010328023694455624, + 0.008268721401691437, + 0.014568474143743515, + 0.017709678038954735, + 0.007868530228734016, + 0.007783127948641777, + 0.010949096642434597, + 0.00741773284971714, + 0.004096062853932381, + 0.0039036159869283438, + 0.006604460068047047, + 0.005076519213616848, + 0.003810113063082099, + 0.004252184182405472, + 0.005294318310916424, + 0.004206178709864616, + 0.014581238850951195, + 0.02003336511552334, + 0.011110733263194561, + 0.011311640962958336, + 0.009285174310207367, + 0.005752976518124342, + 0.008142799139022827, + 0.007262772880494595, + 0.007528060115873814, + 0.010348198935389519, + 0.006216729525476694, + 0.004279640037566423, + 0.004257146269083023, + 0.007134591229259968, + 0.004999092780053616, + 0.005967445205897093, + 0.006603463087230921, + 0.006787536200135946, + 0.007284503895789385, + 0.005249873269349337, + 0.0032476666383445263, + 0.002290805336087942, + 0.004169410094618797, + 0.0034508989192545414, + 0.017131509259343147, + 0.0191792119294405, + 0.012381724081933498, + 0.009013443253934383, + 0.010688385926187038, + 0.014951122924685478, + 0.008125346153974533, + 0.007185983005911112, + 0.008811562322080135, + 0.011822803877294064, + 0.008827866055071354, + 0.007079186383634806, + 0.006962093524634838, + 0.011412122286856174, + 0.011069064028561115, + 0.01364920660853386, + 0.008578528650105, + 0.00720065925270319, + 0.007926926016807556, + 0.006537188310176134, + 0.006641953717917204, + 0.004552445840090513, + 0.007693708874285221, + 0.003890740917995572, + 0.0023039905354380608, + 0.002339398954063654, + 0.0023434064351022243, + 0.0018763112602755427, + 0.002374405739828944, + 0.0017682735342532396, + 0.003229439491406083, + 0.001454220269806683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.16739371418952942 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009811392053961754 + ], + "opposite_logprob_tokens": [ + -4.6348114013671875 + ], + "logprob_reward": -0.009811392053961754, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009811392053961754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20820114016532898, + 0.052513327449560165, + 0.0487983375787735, + 0.03980104997754097, + 0.025692421942949295, + 0.036991145461797714, + 0.13440300524234772, + 0.06292982399463654, + 0.14142407476902008, + 0.24924570322036743 + ], + "credit_map": [ + 0.01814872957766056, + 0.03673696145415306, + 0.022136036306619644, + 0.02460693195462227, + 0.042185187339782715, + 0.019397947937250137, + 0.02887730672955513, + 0.016112016513943672, + 0.00807600375264883, + 0.00922135729342699, + 0.006468795705586672, + 0.006124633830040693, + 0.005339812953025103, + 0.0034175075124949217, + 0.008602233603596687, + 0.005262983031570911, + 0.00632101995870471, + 0.0071525597013533115, + 0.005883903242647648, + 0.007256833836436272, + 0.004944902844727039, + 0.004490351770073175, + 0.008146761916577816, + 0.004602003842592239, + 0.003037870628759265, + 0.004355010576546192, + 0.008687110617756844, + 0.01034370344132185, + 0.0034090508706867695, + 0.0022102040238678455, + 0.0046435873955488205, + 0.0031145112589001656, + 0.0034309441689401865, + 0.002788399113342166, + 0.0034844873007386923, + 0.003234568517655134, + 0.0029766177758574486, + 0.0018354338826611638, + 0.00458942074328661, + 0.0033525521866977215, + 0.0065992893651127815, + 0.0068929921835660934, + 0.006564502138644457, + 0.003951918799430132, + 0.0031578957568854094, + 0.0020064301788806915, + 0.004300085827708244, + 0.003518029348924756, + 0.024611756205558777, + 0.02537606470286846, + 0.016034279018640518, + 0.012826024554669857, + 0.012795649468898773, + 0.018575891852378845, + 0.0115423034876585, + 0.012641034089028835, + 0.007413848303258419, + 0.009409121237695217, + 0.004380847327411175, + 0.00801035389304161, + 0.008964648470282555, + 0.005671932827681303, + 0.009235148318111897, + 0.009843925945460796, + 0.012602061033248901, + 0.025131188333034515, + 0.00874843169003725, + 0.011345110833644867, + 0.028167352080345154, + 0.02515890821814537, + 0.01883002184331417, + 0.011441012844443321, + 0.03295670449733734, + 0.04700503125786781, + 0.06008225679397583, + 0.018547512590885162, + 0.0202507171779871, + 0.03281167522072792, + 0.023209325969219208, + 0.014382487162947655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.009811392053961754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640286788344383 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.004640286788344383, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640286788344383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05294889211654663, + 0.07641003280878067, + 0.05214794725179672, + 0.10285109281539917, + 0.050444189459085464, + 0.07705611735582352, + 0.4103962182998657, + 0.06674297899007797, + 0.024265287443995476, + 0.0867372527718544 + ], + "credit_map": [ + 0.005135949235409498, + 0.005691444966942072, + 0.006914716213941574, + 0.008039052598178387, + 0.00486535532400012, + 0.004776038695126772, + 0.011244751513004303, + 0.006281581707298756, + 0.01084449514746666, + 0.009665342047810555, + 0.009917473420500755, + 0.008205285295844078, + 0.005721703637391329, + 0.009755602106451988, + 0.010451470501720905, + 0.01184866763651371, + 0.007137501146644354, + 0.00454920157790184, + 0.010579921305179596, + 0.005951819475740194, + 0.00397912971675396, + 0.006285833660513163, + 0.005415276624262333, + 0.00824926421046257, + 0.017733003944158554, + 0.017550507560372353, + 0.01753818802535534, + 0.019375713542103767, + 0.005086871329694986, + 0.008034583181142807, + 0.00747430557385087, + 0.01005791686475277, + 0.006299012806266546, + 0.006203711032867432, + 0.008155712857842445, + 0.007477890700101852, + 0.004209174308925867, + 0.007379248738288879, + 0.003967528231441975, + 0.006751908455044031, + 0.01330670528113842, + 0.022234512493014336, + 0.009893501177430153, + 0.0075958105735480785, + 0.004318776074796915, + 0.007259673438966274, + 0.005913965404033661, + 0.006533164996653795, + 0.058746594935655594, + 0.054702457040548325, + 0.05870259180665016, + 0.0426926426589489, + 0.03676682710647583, + 0.07136659324169159, + 0.037216924130916595, + 0.050201594829559326, + 0.007790473289787769, + 0.009909170679748058, + 0.006707708351314068, + 0.0061176191084086895, + 0.005646607372909784, + 0.009682019241154194, + 0.008670351468026638, + 0.012219032272696495, + 0.0028814836405217648, + 0.0031519168987870216, + 0.0025136545300483704, + 0.0029694412369281054, + 0.003721931716427207, + 0.0034233054611831903, + 0.003240290330722928, + 0.0023632629308849573, + 0.011508631519973278, + 0.01824348419904709, + 0.021137408912181854, + 0.006395423319190741, + 0.006786446087062359, + 0.010808411054313183, + 0.007493371609598398, + 0.004364077467471361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.004640286788344383 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639456048607826 + ], + "opposite_logprob_tokens": [ + -5.379639625549316 + ], + "logprob_reward": -0.004639456048607826, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639456048607826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.122355617582798, + 0.051445960998535156, + 0.04782230034470558, + 0.037552956491708755, + 0.07846102863550186, + 0.07659569382667542, + 0.4211044907569885, + 0.03441115468740463, + 0.09603065997362137, + 0.034220173954963684 + ], + "credit_map": [ + 0.010882984846830368, + 0.016790246590971947, + 0.014078543521463871, + 0.014829280786216259, + 0.022301772609353065, + 0.013269414193928242, + 0.020173754543066025, + 0.010029631666839123, + 0.006857849191874266, + 0.0036705886013805866, + 0.006662872154265642, + 0.007272955030202866, + 0.004112846218049526, + 0.007776850834488869, + 0.00709085538983345, + 0.00800115056335926, + 0.007717894855886698, + 0.006049364805221558, + 0.005219880025833845, + 0.00519875343888998, + 0.005667648743838072, + 0.008629568852484226, + 0.004660322330892086, + 0.004678870551288128, + 0.0032293496187776327, + 0.0038452150765806437, + 0.006083384156227112, + 0.006930342875421047, + 0.0036325620021671057, + 0.005645680241286755, + 0.0035306373611092567, + 0.004655790515244007, + 0.009267685003578663, + 0.012012142688035965, + 0.0166276004165411, + 0.01211595255881548, + 0.005198230035603046, + 0.009536609053611755, + 0.006998181808739901, + 0.006704635918140411, + 0.009956722147762775, + 0.01280789915472269, + 0.0094566335901618, + 0.004535998683422804, + 0.011699333786964417, + 0.006906181573867798, + 0.016521435230970383, + 0.004711492452770472, + 0.10291848331689835, + 0.10323914140462875, + 0.05140217766165733, + 0.039258912205696106, + 0.0367262028157711, + 0.041016895323991776, + 0.02622525952756405, + 0.020317435264587402, + 0.006197586189955473, + 0.005232119467109442, + 0.004343060310930014, + 0.004100581631064415, + 0.0028999210335314274, + 0.0033860548865050077, + 0.0037751144263893366, + 0.004476719535887241, + 0.013007336296141148, + 0.013275998644530773, + 0.009328686632215977, + 0.012888068333268166, + 0.012621787376701832, + 0.016506271436810493, + 0.010755524039268494, + 0.0076469918712973595, + 0.004311560653150082, + 0.004784290213137865, + 0.009546911343932152, + 0.0029834702145308256, + 0.0033165377099066973, + 0.003574378089979291, + 0.0036822797264903784, + 0.0020207464694976807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.004639456048607826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160652007907629 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.0160652007907629, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160652007907629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1234886422753334, + 0.07841020077466965, + 0.06901586055755615, + 0.06888880580663681, + 0.09407217055559158, + 0.08312465995550156, + 0.31822335720062256, + 0.045070722699165344, + 0.09657397866249084, + 0.02313159592449665 + ], + "credit_map": [ + 0.012134511955082417, + 0.012067699804902077, + 0.016318682581186295, + 0.017797987908124924, + 0.015788609161973, + 0.009814498014748096, + 0.02376817911863327, + 0.01579846255481243, + 0.0129206208512187, + 0.011817015707492828, + 0.009059050120413303, + 0.011418958194553852, + 0.007819940336048603, + 0.006972568109631538, + 0.010423569940030575, + 0.007978471927344799, + 0.008605951443314552, + 0.008092997595667839, + 0.010449184104800224, + 0.008692054077982903, + 0.007174086757004261, + 0.005082715768367052, + 0.012110104784369469, + 0.008808759041130543, + 0.007443937007337809, + 0.010717601515352726, + 0.013104800134897232, + 0.01126132719218731, + 0.0067537403665483, + 0.004374037962406874, + 0.008187107741832733, + 0.007046252954751253, + 0.011406771838665009, + 0.007274184841662645, + 0.021310755982995033, + 0.016469545662403107, + 0.00850594975054264, + 0.007850018329918385, + 0.009116795845329762, + 0.012138139456510544, + 0.007632868364453316, + 0.015110845677554607, + 0.010013858787715435, + 0.007207322400063276, + 0.010041717439889908, + 0.008702561259269714, + 0.01667141728103161, + 0.0077440571039915085, + 0.06533872336149216, + 0.06699709594249725, + 0.041781265288591385, + 0.029146090149879456, + 0.027521347627043724, + 0.03500603511929512, + 0.02628425508737564, + 0.026148507371544838, + 0.004660159349441528, + 0.006339917425066233, + 0.005644157994538546, + 0.003333136672154069, + 0.005742280278354883, + 0.005967205390334129, + 0.007107377052307129, + 0.00627648551017046, + 0.011718281544744968, + 0.016418812796473503, + 0.009133457206189632, + 0.010266334749758244, + 0.01618545688688755, + 0.010635805316269398, + 0.014097141101956367, + 0.00811868254095316, + 0.002359485486522317, + 0.002647268120199442, + 0.004000133369117975, + 0.003057565540075302, + 0.003503320273011923, + 0.0021181737538427114, + 0.003115489613264799, + 0.002330157905817032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.0160652007907629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, held at eye level, tracks smoothly with the cyclist, then comes to an elegant, static stop as the cyclist comes to a stop. The cyclist, wearing a vibrant red lightweight windbreaker and a sophisticated matte black helmet, rides down a quiet, tree-lined asphalt street. The scene is bathed in the warm, golden light of late afternoon, casting long, soft shadows across the sidewalk. As the rider decides to stop, their pedaling deliberately slows, their leg movements becoming less frequent. The bicycle's wheel gradually slows, the spokes becoming more pronounced with each rotation, visually translating the gentle scraping of the tires on the road. The rider's initially resolute expression, focused on the road ahead, softens noticeably to a calm, reflective one. They bring the bike to a perfectly stable stop, release one foot, and keep it firmly planted on the ground for balance. In this final moment of silence, the tranquil setting, with the golden leaves of the overhanging trees swaying gently in the breeze, creates a dappled light effect that powerfully emphasizes the calm and contemplative atmosphere of the pause.", + "seed": 33430, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_0/index_1029_seed_33430_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05531849320977926, + "vlm_logprob": -0.05531849320977926, + "vlm_yes_no_margin": 3.774999976158142, + "vlm_token_credit": -0.05531849320977926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2165885865688324 + ], + "opposite_logprob_tokens": [ + -2.0915884971618652 + ], + "logprob_reward": -0.2165885865688324, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.2165885865688324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2076186090707779, + 0.09584555774927139, + 0.0870135948061943, + 0.06055909022688866, + 0.25742366909980774, + 0.07435642182826996, + 0.02637886442244053, + 0.04620072618126869, + 0.045442577451467514, + 0.09916094690561295 + ], + "credit_map": [ + 0.014464081265032291, + 0.018140962347388268, + 0.056661300361156464, + 0.019727252423763275, + 0.01563422381877899, + 0.019764430820941925, + 0.020686540752649307, + 0.042539793998003006, + 0.00855844747275114, + 0.010874614119529724, + 0.015107616782188416, + 0.00958382897078991, + 0.008840002119541168, + 0.017331713810563087, + 0.012947706505656242, + 0.012601613067090511, + 0.007003759033977985, + 0.0177844800055027, + 0.010270601138472557, + 0.008494559675455093, + 0.014636027626693249, + 0.010121899656951427, + 0.009419132955372334, + 0.009283135645091534, + 0.005820331629365683, + 0.006192073225975037, + 0.008951451629400253, + 0.0052340091206133366, + 0.005971426144242287, + 0.008374432101845741, + 0.008058754727244377, + 0.011956608854234219, + 0.026688583195209503, + 0.05311453342437744, + 0.027119850739836693, + 0.030053654685616493, + 0.042006541043519974, + 0.03353671357035637, + 0.018642431125044823, + 0.02626132220029831, + 0.0075886985287070274, + 0.005615744274109602, + 0.008698615245521069, + 0.010392808355391026, + 0.013485015369951725, + 0.009729495272040367, + 0.011866431683301926, + 0.0069796061143279076, + 0.003942957613617182, + 0.003619703697040677, + 0.0028005323838442564, + 0.002854752354323864, + 0.00391111196950078, + 0.003234518226236105, + 0.003289035987108946, + 0.002726249396800995, + 0.007861694321036339, + 0.009178685955703259, + 0.005941423587501049, + 0.005991698708385229, + 0.0034152623265981674, + 0.004489013459533453, + 0.004985485691577196, + 0.004337456542998552, + 0.006565785966813564, + 0.006590472999960184, + 0.004233598709106445, + 0.0060412990860641, + 0.00429521594196558, + 0.0041091563180089, + 0.007773201446980238, + 0.005833844188600779, + 0.014926577918231487, + 0.02804802544414997, + 0.012081628665328026, + 0.006775468122214079, + 0.011413577944040298, + 0.011584308929741383, + 0.006681334692984819, + 0.007650019600987434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.2165885865688324 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020626917481422424 + ], + "opposite_logprob_tokens": [ + -3.895627021789551 + ], + "logprob_reward": -0.020626917481422424, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020626917481422424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10713750123977661, + 0.2618476152420044, + 0.050040535628795624, + 0.06437747925519943, + 0.08477520197629929, + 0.04045279696583748, + 0.05537824705243111, + 0.14063197374343872, + 0.03775203600525856, + 0.1576066017150879 + ], + "credit_map": [ + 0.008324330672621727, + 0.010967121459543705, + 0.017866212874650955, + 0.011165229603648186, + 0.009545792825520039, + 0.012116238474845886, + 0.014019835740327835, + 0.023132741451263428, + 0.025228044018149376, + 0.010474234819412231, + 0.08101319521665573, + 0.011049912311136723, + 0.010266164317727089, + 0.021861538290977478, + 0.013744201511144638, + 0.08821031451225281, + 0.004111629445105791, + 0.0070627047680318356, + 0.006336168851703405, + 0.006840444635599852, + 0.007657235488295555, + 0.005064181052148342, + 0.006371020805090666, + 0.006597152445465326, + 0.006668169982731342, + 0.004937127698212862, + 0.009031051769852638, + 0.006087463349103928, + 0.005452411714941263, + 0.014901138842105865, + 0.006895147264003754, + 0.010404969565570354, + 0.009762903675436974, + 0.013677527196705341, + 0.008973967283964157, + 0.0063990275375545025, + 0.013191167265176773, + 0.015949411317706108, + 0.006644309498369694, + 0.01017688773572445, + 0.006457393057644367, + 0.00537294615060091, + 0.004159586038440466, + 0.0034031528048217297, + 0.004562905058264732, + 0.006859879475086927, + 0.003652619430795312, + 0.005984315183013678, + 0.008985462598502636, + 0.007045144215226173, + 0.005124698393046856, + 0.00476776622235775, + 0.007736761122941971, + 0.006268199533224106, + 0.009251685813069344, + 0.006198525428771973, + 0.020890453830361366, + 0.01768524944782257, + 0.018242867663502693, + 0.017233941704034805, + 0.01319402176886797, + 0.021435925737023354, + 0.017958831042051315, + 0.013990684412419796, + 0.0052763596177101135, + 0.006820670794695616, + 0.0033169728703796864, + 0.004810704849660397, + 0.0035525881685316563, + 0.0028149173595011234, + 0.005750652868300676, + 0.005409168545156717, + 0.022593684494495392, + 0.026205357164144516, + 0.012761513702571392, + 0.012469197623431683, + 0.027802160009741783, + 0.027278505265712738, + 0.011779434978961945, + 0.016716742888092995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.020626917481422424 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547007761895657 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012547007761895657, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547007761895657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16107754409313202, + 0.271419495344162, + 0.05584505945444107, + 0.15050199627876282, + 0.03045809268951416, + 0.036870747804641724, + 0.030151238664984703, + 0.07233956456184387, + 0.1356397122144699, + 0.0556965097784996 + ], + "credit_map": [ + 0.017109446227550507, + 0.011461462825536728, + 0.03191502019762993, + 0.015409983694553375, + 0.019143719226121902, + 0.025379188358783722, + 0.01926920749247074, + 0.021389521658420563, + 0.01914350502192974, + 0.00899318978190422, + 0.08923698216676712, + 0.010992191731929779, + 0.012023967690765858, + 0.020390043035149574, + 0.01455539371818304, + 0.09608424454927444, + 0.005826687440276146, + 0.004330217372626066, + 0.008027509786188602, + 0.00395949836820364, + 0.006751164793968201, + 0.012259517796337605, + 0.005463065579533577, + 0.009227405302226543, + 0.014508763328194618, + 0.010123315267264843, + 0.022481519728899002, + 0.01136717852205038, + 0.011778485029935837, + 0.03457798808813095, + 0.015821926295757294, + 0.029842838644981384, + 0.004243929870426655, + 0.0036110945511609316, + 0.003134336555376649, + 0.0028053836431354284, + 0.004387837368994951, + 0.004647756461054087, + 0.004037702456116676, + 0.003590054577216506, + 0.0070849512703716755, + 0.0044209701009094715, + 0.003982361871749163, + 0.004364414140582085, + 0.003974413964897394, + 0.003875813214108348, + 0.00406907731667161, + 0.005098748952150345, + 0.006449930369853973, + 0.005948979873210192, + 0.003799504367634654, + 0.002820635214447975, + 0.002762501360848546, + 0.002392089692875743, + 0.003794889897108078, + 0.0021827092859894037, + 0.017085658386349678, + 0.01410013996064663, + 0.008589034900069237, + 0.006052743643522263, + 0.005925586447119713, + 0.007156641688197851, + 0.0072422693483531475, + 0.006187494844198227, + 0.02218773402273655, + 0.03460387513041496, + 0.019559264183044434, + 0.021655280143022537, + 0.009007562883198261, + 0.008633248507976532, + 0.011628655716776848, + 0.008364097215235233, + 0.007545875385403633, + 0.019496843218803406, + 0.01219846773892641, + 0.00462676677852869, + 0.002915438963100314, + 0.0031213699840009212, + 0.002762484597042203, + 0.003029264509677887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.012547007761895657 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642531931400299 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642531931400299, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642531931400299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19413161277770996, + 0.06525439023971558, + 0.06786263734102249, + 0.11303728818893433, + 0.07955466210842133, + 0.1513483077287674, + 0.03261016681790352, + 0.07254598289728165, + 0.0891474112868309, + 0.13450758159160614 + ], + "credit_map": [ + 0.00886635109782219, + 0.01591726392507553, + 0.06578032672405243, + 0.01570298708975315, + 0.00990118458867073, + 0.010161560960114002, + 0.029427217319607735, + 0.03837471827864647, + 0.0060716005973517895, + 0.006412149406969547, + 0.010121522471308708, + 0.0080167967826128, + 0.006422841921448708, + 0.008851203136146069, + 0.009531325660645962, + 0.009826949797570705, + 0.0058866082690656185, + 0.007854782044887543, + 0.009048445150256157, + 0.0075401924550533295, + 0.009002706035971642, + 0.009969672188162804, + 0.008616109378635883, + 0.00994411762803793, + 0.011139939539134502, + 0.011092943139374256, + 0.01576102152466774, + 0.010831695981323719, + 0.011503616347908974, + 0.021886112168431282, + 0.012396358884871006, + 0.018425604328513145, + 0.00908870343118906, + 0.010274074040353298, + 0.009252116084098816, + 0.005799142178148031, + 0.00926821120083332, + 0.017572127282619476, + 0.00784787442535162, + 0.010452406480908394, + 0.017829280346632004, + 0.02081344835460186, + 0.016276156529784203, + 0.0162043534219265, + 0.01907103694975376, + 0.0269918330013752, + 0.019794756546616554, + 0.014367444440722466, + 0.006015917286276817, + 0.005458666477352381, + 0.004365307744592428, + 0.0035868114791810513, + 0.0029330283869057894, + 0.002426446182653308, + 0.004518832080066204, + 0.0033051567152142525, + 0.014007152989506721, + 0.012892665341496468, + 0.00750025175511837, + 0.007369402330368757, + 0.007085975259542465, + 0.008986535482108593, + 0.007346127647906542, + 0.007357874885201454, + 0.011316250078380108, + 0.015067669562995434, + 0.0068169087171554565, + 0.009048523381352425, + 0.01112682931125164, + 0.010260107927024364, + 0.01314054150134325, + 0.012370576150715351, + 0.018545903265476227, + 0.015596603974699974, + 0.013578478246927261, + 0.007838982157409191, + 0.02302512526512146, + 0.031687699258327484, + 0.01100835483521223, + 0.013226427137851715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.008642531931400299 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187422305345535 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187422305345535, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187422305345535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12578292191028595, + 0.299864798784256, + 0.08688687533140182, + 0.11236894130706787, + 0.08845137059688568, + 0.10074864327907562, + 0.019404252991080284, + 0.060959216207265854, + 0.024072326719760895, + 0.0814606174826622 + ], + "credit_map": [ + 0.009776720777153969, + 0.012202668935060501, + 0.02911650575697422, + 0.020639223977923393, + 0.008675990626215935, + 0.013349873945116997, + 0.01540888287127018, + 0.0166130680590868, + 0.019434623420238495, + 0.014154033735394478, + 0.09627803415060043, + 0.01613238826394081, + 0.011774918995797634, + 0.017922688275575638, + 0.017861677333712578, + 0.10630641877651215, + 0.010170551016926765, + 0.010840940289199352, + 0.010659141466021538, + 0.010737475007772446, + 0.010843080468475819, + 0.011289648711681366, + 0.009137122891843319, + 0.013208913616836071, + 0.013236063532531261, + 0.007920455187559128, + 0.015526941977441311, + 0.008935528807342052, + 0.009806087240576744, + 0.028388245031237602, + 0.010849181562662125, + 0.01770644076168537, + 0.00996573269367218, + 0.011566346511244774, + 0.010227070190012455, + 0.007956322282552719, + 0.010379689745604992, + 0.016671983525156975, + 0.009050204418599606, + 0.01263402123004198, + 0.009511714801192284, + 0.008785302750766277, + 0.011187654919922352, + 0.014254900626838207, + 0.014715554192662239, + 0.013029470108449459, + 0.015900760889053345, + 0.013363287784159184, + 0.003369260346516967, + 0.002083842409774661, + 0.0020394136663526297, + 0.0019739866256713867, + 0.002364562125876546, + 0.0022162683308124542, + 0.0037267059087753296, + 0.0016302147414535284, + 0.009398574009537697, + 0.007943525910377502, + 0.005761006847023964, + 0.0066150035709142685, + 0.00816936045885086, + 0.007938423193991184, + 0.008301048539578915, + 0.00683227414265275, + 0.003798072226345539, + 0.003739687381312251, + 0.003113017650321126, + 0.003419197164475918, + 0.002715446287766099, + 0.002041196683421731, + 0.0027684394735842943, + 0.002477269386872649, + 0.01274430938065052, + 0.006360109429806471, + 0.014349088072776794, + 0.004170258063822985, + 0.012662738561630249, + 0.017177017405629158, + 0.00666819978505373, + 0.007328896317631006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.018187422305345535 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, held at eye level, tracks smoothly with the cyclist, then comes to an elegant, static stop as the cyclist comes to a stop. The cyclist, wearing a vibrant red lightweight windbreaker and a sophisticated matte black helmet, rides down a quiet, tree-lined asphalt street. The scene is bathed in the warm, golden light of late afternoon, casting long, soft shadows across the sidewalk. As the rider decides to stop, their pedaling deliberately slows, their leg movements becoming less frequent. The bicycle's wheel gradually slows, the spokes becoming more pronounced with each rotation, visually translating the gentle scraping of the tires on the road. The rider's initially resolute expression, focused on the road ahead, softens noticeably to a calm, reflective one. They bring the bike to a perfectly stable stop, release one foot, and keep it firmly planted on the ground for balance. In this final moment of silence, the tranquil setting, with the golden leaves of the overhanging trees swaying gently in the breeze, creates a dappled light effect that powerfully emphasizes the calm and contemplative atmosphere of the pause.", + "seed": 33431, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_0/index_1029_seed_33431_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04000485185533762, + "vlm_logprob": -0.04000485185533762, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.04000485185533762, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1601637750864029 + ], + "opposite_logprob_tokens": [ + -2.4101638793945312 + ], + "logprob_reward": -0.1601637750864029, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1601637750864029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08084804564714432, + 0.06362312287092209, + 0.1319059431552887, + 0.09630533307790756, + 0.08695534616708755, + 0.14856866002082825, + 0.19904838502407074, + 0.06355631351470947, + 0.11064774543046951, + 0.018541065976023674 + ], + "credit_map": [ + 0.011723862960934639, + 0.009604097343981266, + 0.0072368113324046135, + 0.007905122824013233, + 0.012929520569741726, + 0.013074123300611973, + 0.008739485405385494, + 0.009635018184781075, + 0.007609044201672077, + 0.008725885301828384, + 0.006265929434448481, + 0.008983723819255829, + 0.012009957805275917, + 0.0035312639083713293, + 0.01062004640698433, + 0.005877272691577673, + 0.013137591071426868, + 0.014531544409692287, + 0.011345998384058475, + 0.015999112278223038, + 0.033218059688806534, + 0.005849178414791822, + 0.026841241866350174, + 0.010983220301568508, + 0.0113532030954957, + 0.014334561303257942, + 0.008610153570771217, + 0.01061191689223051, + 0.007358665112406015, + 0.013427545316517353, + 0.009882871992886066, + 0.02072642184793949, + 0.007389954291284084, + 0.010534814558923244, + 0.011077756062150002, + 0.007944653742015362, + 0.01649470627307892, + 0.008249412290751934, + 0.013635069131851196, + 0.011628983542323112, + 0.012575061991810799, + 0.024416441097855568, + 0.016157513484358788, + 0.010361829772591591, + 0.025287630036473274, + 0.02230176329612732, + 0.01917293108999729, + 0.018295491114258766, + 0.010638725012540817, + 0.057457100600004196, + 0.013201547786593437, + 0.007706027943640947, + 0.013721509836614132, + 0.07051526755094528, + 0.015092210844159126, + 0.010715982876718044, + 0.007241375744342804, + 0.015105151571333408, + 0.005868134554475546, + 0.004831102676689625, + 0.008349861949682236, + 0.009329881519079208, + 0.006022637709975243, + 0.006808162666857243, + 0.009009214118123055, + 0.022108959034085274, + 0.010120931081473827, + 0.01150570623576641, + 0.012814337387681007, + 0.023319197818636894, + 0.011490638367831707, + 0.010278764180839062, + 0.0019426688086241484, + 0.0033853815402835608, + 0.0021726300474256277, + 0.0023095495998859406, + 0.002317033475264907, + 0.0027169554959982634, + 0.002029831986874342, + 0.0016670143231749535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.1601637750864029 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014227032661437988 + ], + "opposite_logprob_tokens": [ + -4.264226913452148 + ], + "logprob_reward": -0.014227032661437988, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014227032661437988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05912426486611366, + 0.023282809183001518, + 0.08511558175086975, + 0.03789237141609192, + 0.09720690548419952, + 0.08969637751579285, + 0.03636498376727104, + 0.02193354442715645, + 0.5339127779006958, + 0.015470299869775772 + ], + "credit_map": [ + 0.007961247116327286, + 0.008320596069097519, + 0.005197342950850725, + 0.007973955012857914, + 0.010657877661287785, + 0.005721823778003454, + 0.006704859435558319, + 0.0065865651704370975, + 0.002903466112911701, + 0.002704014303162694, + 0.002155944472178817, + 0.002884830115363002, + 0.0046989889815449715, + 0.002219403861090541, + 0.0029859421774744987, + 0.00273022148758173, + 0.00775838503614068, + 0.020789775997400284, + 0.022130971774458885, + 0.007977641187608242, + 0.007073513697832823, + 0.0025166822597384453, + 0.010292494669556618, + 0.006576130166649818, + 0.004421605728566647, + 0.004653932061046362, + 0.004443865269422531, + 0.0041461423970758915, + 0.0033762373495846987, + 0.004154894035309553, + 0.005388201214373112, + 0.007307496387511492, + 0.01124656479805708, + 0.01149260625243187, + 0.01189908105880022, + 0.006748037412762642, + 0.017565328627824783, + 0.018672188743948936, + 0.012296069413423538, + 0.007287038955837488, + 0.006629161071032286, + 0.011934751644730568, + 0.011207249015569687, + 0.006639345549046993, + 0.018420103937387466, + 0.010333982296288013, + 0.011949434876441956, + 0.012582351453602314, + 0.002973264316096902, + 0.0061133126728236675, + 0.0035686425399035215, + 0.0025648397859185934, + 0.006279612425714731, + 0.007483309600502253, + 0.004007818177342415, + 0.0033741837833076715, + 0.0016810564557090402, + 0.003809916088357568, + 0.0024030404165387154, + 0.0020177240949124098, + 0.0037041918840259314, + 0.001863015815615654, + 0.0025183632969856262, + 0.003936239052563906, + 0.03269556164741516, + 0.20303697884082794, + 0.019423717632889748, + 0.01867007650434971, + 0.038929857313632965, + 0.16570495069026947, + 0.03382200747728348, + 0.021629707887768745, + 0.0014551709173247218, + 0.0020919216331094503, + 0.002288147108629346, + 0.0013939767377451062, + 0.0033539494033902884, + 0.0016889831749722362, + 0.0016476123128086329, + 0.001550539629533887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.014227032661437988 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036185993812978268 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036185993812978268, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036185993812978268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17998412251472473, + 0.04824173450469971, + 0.116512231528759, + 0.052203986793756485, + 0.16726237535476685, + 0.06618279218673706, + 0.13631142675876617, + 0.10701693594455719, + 0.09692408889532089, + 0.02936028316617012 + ], + "credit_map": [ + 0.02914443053305149, + 0.019381793215870857, + 0.022812262177467346, + 0.015505635179579258, + 0.015866756439208984, + 0.029769113287329674, + 0.012519825249910355, + 0.03498431667685509, + 0.00783691555261612, + 0.005276598036289215, + 0.00418237317353487, + 0.004811588209122419, + 0.004946303088217974, + 0.006774716544896364, + 0.005927640944719315, + 0.008485603146255016, + 0.0172297190874815, + 0.016165776178240776, + 0.020389549434185028, + 0.012468939647078514, + 0.017908640205860138, + 0.007853599265217781, + 0.015195364132523537, + 0.009300652891397476, + 0.007506170310080051, + 0.006114959716796875, + 0.006226928438991308, + 0.007803683634847403, + 0.005046954844146967, + 0.004705369006842375, + 0.0050029815174639225, + 0.009796942584216595, + 0.018725359812378883, + 0.01889275573194027, + 0.018396692350506783, + 0.011932743713259697, + 0.031951162964105606, + 0.032699842005968094, + 0.022504102438688278, + 0.0121597396209836, + 0.006114968564361334, + 0.010056392289698124, + 0.008399752900004387, + 0.004456925205886364, + 0.009007672779262066, + 0.008934170007705688, + 0.008453430607914925, + 0.010759478434920311, + 0.008556085638701916, + 0.034241121262311935, + 0.010715642012655735, + 0.006602445617318153, + 0.011637275107204914, + 0.04682208225131035, + 0.01042064931243658, + 0.007316131144762039, + 0.011763334274291992, + 0.02497553825378418, + 0.009802920743823051, + 0.012501029297709465, + 0.009807942435145378, + 0.013714182190597057, + 0.009732721373438835, + 0.014719273895025253, + 0.007728131953626871, + 0.02504657581448555, + 0.007927874103188515, + 0.006386158987879753, + 0.011517050676047802, + 0.01943856105208397, + 0.009223723784089088, + 0.009656019508838654, + 0.0029043315444141626, + 0.006093979347497225, + 0.004682883620262146, + 0.002745360601693392, + 0.0029187030158936977, + 0.003519516671076417, + 0.0030691134743392467, + 0.003426398616284132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0036185993812978268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950355436652899 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950355436652899, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950355436652899, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08334609866142273, + 0.06369445472955704, + 0.1291031539440155, + 0.09559434652328491, + 0.08505143970251083, + 0.08980731666088104, + 0.18760880827903748, + 0.08085361868143082, + 0.15626657009124756, + 0.028674181550741196 + ], + "credit_map": [ + 0.01007350068539381, + 0.008621499873697758, + 0.008866791613399982, + 0.012990309856832027, + 0.014673170633614063, + 0.0066204736940562725, + 0.009977337904274464, + 0.011523011140525341, + 0.008053142577409744, + 0.007835205644369125, + 0.005769472569227219, + 0.005778506398200989, + 0.006078755483031273, + 0.009837985038757324, + 0.007479348219931126, + 0.012862029485404491, + 0.010598977096378803, + 0.01778080314397812, + 0.016720088198781013, + 0.012413832359015942, + 0.024196116253733635, + 0.007815294899046421, + 0.028135087341070175, + 0.01144295185804367, + 0.01388803031295538, + 0.010352731682360172, + 0.01120047178119421, + 0.010458066128194332, + 0.009654096327722073, + 0.00973687507212162, + 0.01419273391366005, + 0.016111336648464203, + 0.009630079381167889, + 0.015139802359044552, + 0.008983013220131397, + 0.008508454076945782, + 0.011330785229802132, + 0.012317148968577385, + 0.01217591855674982, + 0.006966236047446728, + 0.010083531960844994, + 0.013276736252009869, + 0.009887580759823322, + 0.007562472019344568, + 0.013940460048615932, + 0.009980241768062115, + 0.014165224507451057, + 0.010911054909229279, + 0.009545248933136463, + 0.055050913244485855, + 0.00961423572152853, + 0.007282918784767389, + 0.015926044434309006, + 0.06648845225572586, + 0.014752493239939213, + 0.008948486298322678, + 0.008449858985841274, + 0.01703522354364395, + 0.007439972832798958, + 0.007401823997497559, + 0.016484268009662628, + 0.007038783747702837, + 0.009989862330257893, + 0.007013821043074131, + 0.011118749156594276, + 0.036215655505657196, + 0.012805587612092495, + 0.013066282495856285, + 0.015365839004516602, + 0.04232660308480263, + 0.015047106891870499, + 0.01032074261456728, + 0.0030617965385317802, + 0.0047827111557126045, + 0.006708494853228331, + 0.002736755646765232, + 0.0024694453459233046, + 0.0033379283268004656, + 0.00319185433909297, + 0.0023851937148720026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.005950355436652899 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot held at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cyclist wearing a vibrant red lightweight windbreaker and a matte black helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene take place on a quiet, tree-lined asphalt street bathed in warm golden light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera track smoothly with the cyclist and then come to a static stop as the cyclist stops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the cyclist's pedaling slowing down and the wheel spokes becoming more pronounced as they stop?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064496710896492 + ], + "opposite_logprob_tokens": [ + -4.141064643859863 + ], + "logprob_reward": -0.016064496710896492, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064496710896492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1181156113743782, + 0.0403580516576767, + 0.3282849192619324, + 0.09829708188772202, + 0.0704847052693367, + 0.0853617936372757, + 0.04449693486094475, + 0.03476853296160698, + 0.15309181809425354, + 0.02674056589603424 + ], + "credit_map": [ + 0.016493981704115868, + 0.014072274789214134, + 0.013774504885077477, + 0.013030684553086758, + 0.013913622125983238, + 0.010642029345035553, + 0.015274913050234318, + 0.020913612097501755, + 0.004950289614498615, + 0.0033427018206566572, + 0.003775026649236679, + 0.003952073864638805, + 0.007125838194042444, + 0.00447682524099946, + 0.007593913469463587, + 0.005141384433954954, + 0.024878496304154396, + 0.04021550342440605, + 0.05653107911348343, + 0.04070886969566345, + 0.06761528551578522, + 0.015672530978918076, + 0.06469345837831497, + 0.01796969212591648, + 0.009903441183269024, + 0.018809335306286812, + 0.009893344715237617, + 0.010050270706415176, + 0.01045385841280222, + 0.009253023192286491, + 0.012304553762078285, + 0.01762925460934639, + 0.009059039875864983, + 0.0076994821429252625, + 0.009190109558403492, + 0.005646813195198774, + 0.010485180653631687, + 0.012511196546256542, + 0.010332350619137287, + 0.005560539662837982, + 0.0064095607958734035, + 0.012486286461353302, + 0.008267747238278389, + 0.006211576517671347, + 0.01933397352695465, + 0.010791883803904057, + 0.010913804173469543, + 0.010946960188448429, + 0.003807970555499196, + 0.009668341837823391, + 0.0029499458614736795, + 0.002259977860376239, + 0.0038290079683065414, + 0.014107611030340195, + 0.004581007175147533, + 0.0032930749002844095, + 0.002953245071694255, + 0.005845499224960804, + 0.004663341213017702, + 0.003984897397458553, + 0.006931052077561617, + 0.0031808752100914717, + 0.00419040210545063, + 0.003019218798726797, + 0.009592477232217789, + 0.04475589469075203, + 0.011759774759411812, + 0.009849941357970238, + 0.012812461704015732, + 0.03858399763703346, + 0.013773900456726551, + 0.01196337305009365, + 0.0027531671803444624, + 0.004397390875965357, + 0.00676013482734561, + 0.0017962065758183599, + 0.002832717029377818, + 0.002903550863265991, + 0.002806540811434388, + 0.0024908571504056454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.016064496710896492 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_1/index_tensor([2484, 2484])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_1/index_tensor([2484, 2484])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e25e56b35a62cd55a83186989e92f28f621e6f03 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_1/index_tensor([2484, 2484])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Captured with the precision of ultra-high-speed photography, this extreme macro close-up slowly traces the richly textured, deep green surface of a rare spiny bamboo stalk. Other bamboo stems blur in the background, while soft golden morning light filters through the treetops, creating a dreamy, verdant bokeh effect and casting a delicate rim light across the scene. The camera's focus falls on a cluster of small, tightly sealed flower buds emerging from a spiny node on the stalk. In breathtaking, ultra-slow detail, one bud begins to tremble and swell, its outer skin splitting open in an elegant, explosive manner, revealing the intricate structure within. The delicate, paper-thin, cream-colored florets unfold in a fluid, almost liquid motion. Simultaneously, a cluster of thin, thread-like stamens burst forth, their pollen releasing a shimmering cloud of tiny golden pollen that hangs fleetingly in the still air. The camera meticulously tracks this rapid, graceful eruption, transforming the complex movements, normally invisible to the naked eye, into a spectacle of natural art, evoking a profound awe and the quiet, powerful beauty of a rare natural phenomenon.", + "seed": 15244, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_1/index_2484_seed_15244_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.007292566681280732, + "vlm_logprob": -0.007292566681280732, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.007292566681280732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01339601818472147 + ], + "opposite_logprob_tokens": [ + -4.6383957862854 + ], + "logprob_reward": -0.01339601818472147, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01339601818472147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11945789307355881, + 0.1458333432674408, + 0.05546656623482704, + 0.13271594047546387, + 0.12596483528614044, + 0.07172904908657074, + 0.05783378705382347, + 0.12293399125337601, + 0.09432566910982132, + 0.0737389326095581 + ], + "credit_map": [ + 0.022357584908604622, + 0.010986930690705776, + 0.020765531808137894, + 0.017868245020508766, + 0.016637522727251053, + 0.011611740104854107, + 0.008764808066189289, + 0.010465523228049278, + 0.030135920271277428, + 0.01083100214600563, + 0.01846855878829956, + 0.025639189407229424, + 0.01945699192583561, + 0.018044153228402138, + 0.011682425625622272, + 0.011575098149478436, + 0.010042703710496426, + 0.006282810587435961, + 0.007882007397711277, + 0.006108403671532869, + 0.007009974680840969, + 0.006624284666031599, + 0.005274479277431965, + 0.006241899449378252, + 0.017363643273711205, + 0.014051291160285473, + 0.019560571759939194, + 0.018003681674599648, + 0.02698843739926815, + 0.011068173684179783, + 0.012282022275030613, + 0.013398118317127228, + 0.02783992514014244, + 0.013970092870295048, + 0.018768107518553734, + 0.01833071932196617, + 0.012887275777757168, + 0.013780881650745869, + 0.010160953737795353, + 0.010226878337562084, + 0.01088674459606409, + 0.01301948819309473, + 0.009271283634006977, + 0.00866717379540205, + 0.006482360884547234, + 0.006275984924286604, + 0.006371438968926668, + 0.01075457502156496, + 0.011846065521240234, + 0.014284895732998848, + 0.008371684700250626, + 0.006006050854921341, + 0.0046159978955984116, + 0.005314534064382315, + 0.003481350839138031, + 0.003913207445293665, + 0.020386720076203346, + 0.028528369963169098, + 0.017410367727279663, + 0.01501783262938261, + 0.013703763484954834, + 0.00881592184305191, + 0.010314173065125942, + 0.008756840601563454, + 0.013982020318508148, + 0.014443167485296726, + 0.012251549400389194, + 0.010676998645067215, + 0.014474544674158096, + 0.008539156056940556, + 0.011188933625817299, + 0.008769296109676361, + 0.013022655621170998, + 0.010594848543405533, + 0.009151959791779518, + 0.008777274750173092, + 0.008587900549173355, + 0.005644150543957949, + 0.008197233080863953, + 0.009762914851307869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.01339601818472147 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002829835517331958 + ], + "opposite_logprob_tokens": [ + -5.8778300285339355 + ], + "logprob_reward": -0.002829835517331958, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002829835517331958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10572122782468796, + 0.07794801890850067, + 0.2792215347290039, + 0.06595878303050995, + 0.036720581352710724, + 0.06614851206541061, + 0.06994763016700745, + 0.08123084157705307, + 0.09149318188428879, + 0.12560978531837463 + ], + "credit_map": [ + 0.0199537742882967, + 0.011801234446465969, + 0.010313896462321281, + 0.02496037632226944, + 0.013250531628727913, + 0.006803571712225676, + 0.009965169243514538, + 0.008672671392560005, + 0.012918381951749325, + 0.010254713706672192, + 0.008071289397776127, + 0.016401061788201332, + 0.012321112677454948, + 0.005625081714242697, + 0.00582627858966589, + 0.006530097685754299, + 0.05233439803123474, + 0.02247423306107521, + 0.0408792681992054, + 0.023229019716382027, + 0.035367026925086975, + 0.041368987411260605, + 0.026726707816123962, + 0.03684186190366745, + 0.011050085537135601, + 0.006073995027691126, + 0.010109403170645237, + 0.009544968605041504, + 0.011946548707783222, + 0.005283166654407978, + 0.0055271633900702, + 0.0064234500750899315, + 0.008029558695852757, + 0.00456397095695138, + 0.005695714615285397, + 0.0035599302500486374, + 0.002833317033946514, + 0.004395990166813135, + 0.003750767558813095, + 0.0038913311436772346, + 0.007928087376058102, + 0.014562739059329033, + 0.007777429651468992, + 0.00843765027821064, + 0.00716091925278306, + 0.006098002661019564, + 0.0051528168842196465, + 0.009030863642692566, + 0.010085228830575943, + 0.01591983251273632, + 0.009886796586215496, + 0.00628232816234231, + 0.008743910118937492, + 0.005192723590880632, + 0.004678942728787661, + 0.009157869033515453, + 0.0148925194516778, + 0.014657821506261826, + 0.010952482931315899, + 0.013930597342550755, + 0.006341622676700354, + 0.005266325082629919, + 0.007301692385226488, + 0.007887777872383595, + 0.015252649784088135, + 0.015423096716403961, + 0.012824894860386848, + 0.010380133055150509, + 0.012769430875778198, + 0.007487793918699026, + 0.009790117852389812, + 0.007565057836472988, + 0.020488163456320763, + 0.02214270643889904, + 0.01431218534708023, + 0.012665793299674988, + 0.020676083862781525, + 0.009920421056449413, + 0.011140521615743637, + 0.014263907447457314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.002829835517331958 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028271013870835304 + ], + "opposite_logprob_tokens": [ + -5.877827167510986 + ], + "logprob_reward": -0.0028271013870835304, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028271013870835304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07810576260089874, + 0.11860796809196472, + 0.07893014699220657, + 0.06190080568194389, + 0.047886185348033905, + 0.08017405867576599, + 0.09597528725862503, + 0.06680267304182053, + 0.1069917231798172, + 0.26462531089782715 + ], + "credit_map": [ + 0.010852299630641937, + 0.011020987294614315, + 0.009154774248600006, + 0.013101448304951191, + 0.012514322996139526, + 0.00817156583070755, + 0.0077986461110413074, + 0.005491721443831921, + 0.018355706706643105, + 0.014974410645663738, + 0.010401494801044464, + 0.02461926080286503, + 0.017862018197774887, + 0.009811939671635628, + 0.01119096390902996, + 0.011392188258469105, + 0.011519894935190678, + 0.011181844398379326, + 0.011730009689927101, + 0.014582939445972443, + 0.01164124347269535, + 0.0056683700531721115, + 0.006369016598910093, + 0.006236840970814228, + 0.010334216058254242, + 0.0057867225259542465, + 0.008496630005538464, + 0.009057780727744102, + 0.010461606085300446, + 0.005108021665364504, + 0.005443946458399296, + 0.007211894262582064, + 0.010740162804722786, + 0.0071265422739088535, + 0.006726458203047514, + 0.006483832374215126, + 0.004489914048463106, + 0.00557232741266489, + 0.0035390492994338274, + 0.003207903355360031, + 0.008795144036412239, + 0.019238831475377083, + 0.008664102293550968, + 0.011079240590333939, + 0.009507226757705212, + 0.007025319617241621, + 0.006078919395804405, + 0.009785283356904984, + 0.012199955061078072, + 0.024197300896048546, + 0.01308504305779934, + 0.009735975414514542, + 0.01248879637569189, + 0.008203021250665188, + 0.008790218271315098, + 0.007274982053786516, + 0.013011740520596504, + 0.011997348628938198, + 0.011135554872453213, + 0.006238587666302919, + 0.00730694318190217, + 0.004503560718148947, + 0.0053488388657569885, + 0.007260109297931194, + 0.01474117673933506, + 0.014856030233204365, + 0.012590194121003151, + 0.012008016929030418, + 0.01800275221467018, + 0.013145234435796738, + 0.012913212180137634, + 0.008735126815736294, + 0.0395965576171875, + 0.0522143617272377, + 0.027895037084817886, + 0.027006980031728745, + 0.04512741416692734, + 0.02314629964530468, + 0.023564869537949562, + 0.026073843240737915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.0028271013870835304 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784241206943989 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784241206943989, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784241206943989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03435169905424118, + 0.018268557265400887, + 0.024214688688516617, + 0.026931539177894592, + 0.013109421357512474, + 0.05420035868883133, + 0.028008202090859413, + 0.02691151201725006, + 0.09817884117364883, + 0.6758251786231995 + ], + "credit_map": [ + 0.007218161132186651, + 0.0030050380155444145, + 0.005325390491634607, + 0.007778698578476906, + 0.003024172270670533, + 0.0026807948015630245, + 0.0024732600431889296, + 0.0028461827896535397, + 0.002875172533094883, + 0.002401873003691435, + 0.001874152454547584, + 0.0029447015840560198, + 0.002397209871560335, + 0.0018063911702483892, + 0.001879907213151455, + 0.0020891493186354637, + 0.004585602320730686, + 0.002265611430630088, + 0.003873092122375965, + 0.0021133276168257, + 0.003116470528766513, + 0.003422774374485016, + 0.002284543588757515, + 0.0025532683357596397, + 0.004194202832877636, + 0.0028383443132042885, + 0.004033857025206089, + 0.004092380404472351, + 0.004990183282643557, + 0.002434487920254469, + 0.002318419748917222, + 0.0020296655129641294, + 0.002132580615580082, + 0.001645806827582419, + 0.001736765494570136, + 0.0018643540097400546, + 0.001361218048259616, + 0.002084798878058791, + 0.0010258639231324196, + 0.001258032862097025, + 0.006307491101324558, + 0.01236667949706316, + 0.006242032628506422, + 0.007786106783896685, + 0.006056095473468304, + 0.00479521369561553, + 0.0038402036298066378, + 0.006806532386690378, + 0.004144127946346998, + 0.007337913382798433, + 0.004409415181726217, + 0.003034143941476941, + 0.0032265419140458107, + 0.0017430144362151623, + 0.001944545074366033, + 0.00216849846765399, + 0.004085520748049021, + 0.005794372409582138, + 0.0038924154359847307, + 0.003148798132315278, + 0.002452184446156025, + 0.0021471732761710882, + 0.002595870988443494, + 0.0027951744850724936, + 0.017411433160305023, + 0.014894717372953892, + 0.012659854255616665, + 0.009020514786243439, + 0.019250303506851196, + 0.010859164409339428, + 0.00849694013595581, + 0.005585914012044668, + 0.10081465542316437, + 0.13618431985378265, + 0.07083150744438171, + 0.07087922841310501, + 0.1240939199924469, + 0.0564030222594738, + 0.058010440319776535, + 0.058608103543519974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.009784241206943989 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625637110322714 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625637110322714, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625637110322714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05359738692641258, + 0.10904447734355927, + 0.07446540147066116, + 0.0687774196267128, + 0.08677376806735992, + 0.25733792781829834, + 0.02870871126651764, + 0.05337138473987579, + 0.11586974561214447, + 0.15205377340316772 + ], + "credit_map": [ + 0.011132854036986828, + 0.00550789013504982, + 0.005668870639055967, + 0.01004951260983944, + 0.0045880964025855064, + 0.0050997366197407246, + 0.004807665012776852, + 0.0067427619360387325, + 0.019903777167201042, + 0.008096931502223015, + 0.016931505873799324, + 0.018421076238155365, + 0.013293737545609474, + 0.013311388902366161, + 0.009777248837053776, + 0.009308816865086555, + 0.017562387511134148, + 0.007214141543954611, + 0.010788329876959324, + 0.007775569334626198, + 0.00930399727076292, + 0.006559466943144798, + 0.006657950114458799, + 0.008603558875620365, + 0.010745301842689514, + 0.0068947128020226955, + 0.011038435623049736, + 0.009931649081408978, + 0.01227777823805809, + 0.006095719989389181, + 0.006267355289310217, + 0.005526468623429537, + 0.015785139054059982, + 0.013613568618893623, + 0.012032958678901196, + 0.013908839784562588, + 0.010231207124888897, + 0.008090570569038391, + 0.005959867499768734, + 0.007151616737246513, + 0.03479776158928871, + 0.0554472953081131, + 0.030355246737599373, + 0.0329265333712101, + 0.02710072696208954, + 0.026513313874602318, + 0.01900186948478222, + 0.03119518980383873, + 0.005091810133308172, + 0.005628680344671011, + 0.004979873076081276, + 0.0026233328972011805, + 0.002726728795096278, + 0.0030280654318630695, + 0.0018929033540189266, + 0.0027373163029551506, + 0.00955495610833168, + 0.013015005737543106, + 0.009808080270886421, + 0.005215357057750225, + 0.00471162423491478, + 0.004074534866958857, + 0.0035019030328840017, + 0.003489925991743803, + 0.015370677225291729, + 0.017168300226330757, + 0.012969538569450378, + 0.009832872077822685, + 0.02310357056558132, + 0.013827111572027206, + 0.01442462857812643, + 0.009173046797513962, + 0.021437890827655792, + 0.031702250242233276, + 0.017706258222460747, + 0.015989739447832108, + 0.028717301785945892, + 0.012278365902602673, + 0.015207473188638687, + 0.009014485403895378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.007625637110322714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured with the precision of ultra-high-speed photography, this extreme macro close-up slowly traces the richly textured, deep green surface of a rare spiny bamboo stalk. Other bamboo stems blur in the background, while soft golden morning light filters through the treetops, creating a dreamy, verdant bokeh effect and casting a delicate rim light across the scene. The camera's focus falls on a cluster of small, tightly sealed flower buds emerging from a spiny node on the stalk. In breathtaking, ultra-slow detail, one bud begins to tremble and swell, its outer skin splitting open in an elegant, explosive manner, revealing the intricate structure within. The delicate, paper-thin, cream-colored florets unfold in a fluid, almost liquid motion. Simultaneously, a cluster of thin, thread-like stamens burst forth, their pollen releasing a shimmering cloud of tiny golden pollen that hangs fleetingly in the still air. The camera meticulously tracks this rapid, graceful eruption, transforming the complex movements, normally invisible to the naked eye, into a spectacle of natural art, evoking a profound awe and the quiet, powerful beauty of a rare natural phenomenon.", + "seed": 15245, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_1/index_2484_seed_15245_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.005668399785645306, + "vlm_logprob": -0.005668399785645306, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.005668399785645306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008942436426877975 + ], + "opposite_logprob_tokens": [ + -5.008942604064941 + ], + "logprob_reward": -0.008942436426877975, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008942436426877975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1407882422208786, + 0.12073294818401337, + 0.16155624389648438, + 0.08290316164493561, + 0.03397109732031822, + 0.26002833247184753, + 0.039616908878088, + 0.05516398698091507, + 0.0763283222913742, + 0.02891075238585472 + ], + "credit_map": [ + 0.015173320658504963, + 0.011943901889026165, + 0.015972578898072243, + 0.018464766442775726, + 0.013766433112323284, + 0.024075457826256752, + 0.02397295832633972, + 0.017418820410966873, + 0.01290062628686428, + 0.007444795686751604, + 0.03204530477523804, + 0.011041251011192799, + 0.00899918470531702, + 0.01052903849631548, + 0.009403684176504612, + 0.028369057923555374, + 0.0176707673817873, + 0.006265563424676657, + 0.05651713162660599, + 0.009849227033555508, + 0.006635166238993406, + 0.013733030296862125, + 0.008421918377280235, + 0.042463455349206924, + 0.009130953811109066, + 0.005775758530944586, + 0.026214104145765305, + 0.006660824175924063, + 0.005292167887091637, + 0.006623236928135157, + 0.005293946713209152, + 0.01791217178106308, + 0.0048497323878109455, + 0.0029897091444581747, + 0.004900916013866663, + 0.0032387205865234137, + 0.005015615373849869, + 0.005213049240410328, + 0.003971110098063946, + 0.003792245639488101, + 0.025227811187505722, + 0.011236114427447319, + 0.0848623588681221, + 0.018481582403182983, + 0.01026797667145729, + 0.014904187992215157, + 0.020684044808149338, + 0.07436424493789673, + 0.004957288503646851, + 0.005730308126658201, + 0.008186518214643002, + 0.002750847954303026, + 0.0035475012846291065, + 0.004134526476264, + 0.0034452511463314295, + 0.006864667404443026, + 0.008136799558997154, + 0.007773967459797859, + 0.007247625384479761, + 0.005296472925692797, + 0.008120319806039333, + 0.005999737419188023, + 0.005132483784109354, + 0.007456580176949501, + 0.00832926295697689, + 0.012827163562178612, + 0.009517417289316654, + 0.008194615133106709, + 0.011636373586952686, + 0.007551093120127916, + 0.007615262642502785, + 0.010657135397195816, + 0.0033024679869413376, + 0.003713285317644477, + 0.003753169672563672, + 0.003541254438459873, + 0.0023985514417290688, + 0.00362138613127172, + 0.003949959762394428, + 0.004630676005035639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.008942436426877975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017203782917931676 + ], + "opposite_logprob_tokens": [ + -6.376720428466797 + ], + "logprob_reward": -0.0017203782917931676, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017203782917931676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07552500814199448, + 0.09720484912395477, + 0.08230150490999222, + 0.10573805868625641, + 0.059345122426748276, + 0.39247816801071167, + 0.02986505627632141, + 0.08773592114448547, + 0.03706732764840126, + 0.032738979905843735 + ], + "credit_map": [ + 0.009355958551168442, + 0.0064269741997122765, + 0.009013799019157887, + 0.010683842934668064, + 0.010094367899000645, + 0.009327470324933529, + 0.010400033555924892, + 0.010222560726106167, + 0.009313060902059078, + 0.0051811919547617435, + 0.025045324116945267, + 0.008070864714682102, + 0.007623948156833649, + 0.010403386317193508, + 0.008217178285121918, + 0.02334989234805107, + 0.008393781259655952, + 0.006230467930436134, + 0.018026813864707947, + 0.009662180207669735, + 0.0055479384027421474, + 0.012152097187936306, + 0.007633662782609463, + 0.014654558151960373, + 0.010656195692718029, + 0.016405602917075157, + 0.01518154051154852, + 0.015349002555012703, + 0.012825356796383858, + 0.014209149405360222, + 0.010585596784949303, + 0.010525617748498917, + 0.005332242697477341, + 0.006650309544056654, + 0.01110390666872263, + 0.007716601714491844, + 0.0075516775250434875, + 0.006366093177348375, + 0.0058659035712480545, + 0.008758388459682465, + 0.04486222192645073, + 0.020749753341078758, + 0.11904453486204147, + 0.018838996067643166, + 0.020662259310483932, + 0.028611620888113976, + 0.027098238468170166, + 0.1126105859875679, + 0.005222166422754526, + 0.004900467116385698, + 0.0040756771340966225, + 0.0025325841270387173, + 0.0032215826213359833, + 0.0031386730261147022, + 0.0022742380388081074, + 0.004499668721109629, + 0.00863014068454504, + 0.009184843860566616, + 0.012839122675359249, + 0.011489336378872395, + 0.01185217872262001, + 0.010188961401581764, + 0.008519238792359829, + 0.015032103285193443, + 0.0041974205523729324, + 0.004202567506581545, + 0.006181600969284773, + 0.0049949404783546925, + 0.003754633478820324, + 0.0038704616017639637, + 0.004218809772282839, + 0.005646895617246628, + 0.0035142600536346436, + 0.004412952810525894, + 0.004600117914378643, + 0.0034698299132287502, + 0.0033898530527949333, + 0.004434848669916391, + 0.004160378128290176, + 0.004756740294396877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.0017203782917931676 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956517532467842 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956517532467842, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956517532467842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07343747466802597, + 0.06316982954740524, + 0.1662377417087555, + 0.054141655564308167, + 0.10712046921253204, + 0.11385735869407654, + 0.08806949853897095, + 0.0874357521533966, + 0.20470215380191803, + 0.04182807356119156 + ], + "credit_map": [ + 0.008030716329813004, + 0.004722817335277796, + 0.010314952582120895, + 0.007813097909092903, + 0.008418230339884758, + 0.012377730570733547, + 0.013119316659867764, + 0.008640604093670845, + 0.010751035064458847, + 0.0028855307027697563, + 0.01960170827805996, + 0.004578453954309225, + 0.004288839641958475, + 0.004587571602314711, + 0.0035880515351891518, + 0.012888645753264427, + 0.02114340290427208, + 0.009409599006175995, + 0.0503162145614624, + 0.012770186178386211, + 0.009030812419950962, + 0.01622680388391018, + 0.012857040390372276, + 0.03448367863893509, + 0.0058443848975002766, + 0.006412457674741745, + 0.008362538181245327, + 0.007818379439413548, + 0.008625218644738197, + 0.006327167619019747, + 0.0058524757623672485, + 0.0048990328796207905, + 0.009990175254642963, + 0.01171619351953268, + 0.01660170592367649, + 0.020133862271904945, + 0.01305472943931818, + 0.009314512833952904, + 0.012228225357830524, + 0.014081059955060482, + 0.010634480975568295, + 0.010476273484528065, + 0.02633122354745865, + 0.016854699701070786, + 0.009218615479767323, + 0.008237387053668499, + 0.01184334047138691, + 0.020261336117982864, + 0.012165762484073639, + 0.004456823691725731, + 0.024148233234882355, + 0.00534470472484827, + 0.006951695773750544, + 0.007917417213320732, + 0.006674997042864561, + 0.02040986903011799, + 0.007068476174026728, + 0.013805285096168518, + 0.011847337707877159, + 0.01424364373087883, + 0.00865780096501112, + 0.00798860564827919, + 0.008740220218896866, + 0.015084386803209782, + 0.026919037103652954, + 0.030141262337565422, + 0.01843298226594925, + 0.024232765659689903, + 0.03728186711668968, + 0.018448514863848686, + 0.021394193172454834, + 0.027851536870002747, + 0.004555709194391966, + 0.005953569896519184, + 0.006226111203432083, + 0.004781064111739397, + 0.0033454326912760735, + 0.005826176609843969, + 0.004743196070194244, + 0.006396811455488205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.005956517532467842 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629659026861191 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629659026861191, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629659026861191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057986024767160416, + 0.07818155735731125, + 0.08942407369613647, + 0.08526454120874405, + 0.03653419017791748, + 0.18607328832149506, + 0.03493012487888336, + 0.10682748258113861, + 0.2986609935760498, + 0.026117749512195587 + ], + "credit_map": [ + 0.005837524309754372, + 0.004373406060039997, + 0.00901800673455, + 0.008618512190878391, + 0.0063242604956030846, + 0.00677859503775835, + 0.0069136549718678, + 0.010122067295014858, + 0.007860484533011913, + 0.00896908063441515, + 0.012544717639684677, + 0.010989222675561905, + 0.007501049432903528, + 0.010929334908723831, + 0.010134972631931305, + 0.009252692572772503, + 0.010592171922326088, + 0.005644451826810837, + 0.016072457656264305, + 0.010398253798484802, + 0.006210125982761383, + 0.016520438715815544, + 0.01124651450663805, + 0.012739651836454868, + 0.00858344230800867, + 0.00515721132978797, + 0.02572677843272686, + 0.008379141800105572, + 0.004778207279741764, + 0.007273741532117128, + 0.006428578868508339, + 0.01893744431436062, + 0.0035956655628979206, + 0.003995752427726984, + 0.004909059964120388, + 0.005789404269307852, + 0.004182473756372929, + 0.005906122270971537, + 0.003886438673362136, + 0.004269270226359367, + 0.01686283014714718, + 0.006980573292821646, + 0.061937350779771805, + 0.012036347761750221, + 0.006139443721622229, + 0.011218457482755184, + 0.015845883637666702, + 0.05505240336060524, + 0.004670143127441406, + 0.004786753095686436, + 0.0048458632081747055, + 0.00336475926451385, + 0.004327950533479452, + 0.004018588922917843, + 0.004082983359694481, + 0.004833084065467119, + 0.009266313165426254, + 0.014328976161777973, + 0.014453676529228687, + 0.017036374658346176, + 0.011545016430318356, + 0.011277671903371811, + 0.011460025794804096, + 0.017459433525800705, + 0.025530386716127396, + 0.03812003135681152, + 0.03630233183503151, + 0.049800172448158264, + 0.0347578302025795, + 0.028892209753394127, + 0.03668947517871857, + 0.04856855049729347, + 0.003304137848317623, + 0.0032390078995376825, + 0.003638568567112088, + 0.002995403716340661, + 0.0023679451551288366, + 0.003281960729509592, + 0.0037569506093859673, + 0.0035337749868631363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.007629659026861191 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093007650226355 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093007650226355, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093007650226355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2432524412870407, + 0.1523979753255844, + 0.09759312868118286, + 0.0447341650724411, + 0.0691150650382042, + 0.10133900493383408, + 0.03561289235949516, + 0.13284191489219666, + 0.0852520763874054, + 0.037861306220293045 + ], + "credit_map": [ + 0.03216329589486122, + 0.018048452213406563, + 0.05051160603761673, + 0.02427959442138672, + 0.020872974768280983, + 0.023756781592965126, + 0.02633562870323658, + 0.047284085303545, + 0.01750444434583187, + 0.007475555874407291, + 0.048120975494384766, + 0.012508646585047245, + 0.007906369864940643, + 0.007625443395227194, + 0.009215440601110458, + 0.0420411080121994, + 0.008349205367267132, + 0.005508312955498695, + 0.024463610723614693, + 0.008620770648121834, + 0.006221422925591469, + 0.015749800950288773, + 0.009352872148156166, + 0.019327130168676376, + 0.004861763678491116, + 0.005427794065326452, + 0.006056115962564945, + 0.005655727349221706, + 0.004150809720158577, + 0.007662074640393257, + 0.006015918217599392, + 0.004903961438685656, + 0.007866275496780872, + 0.003845562692731619, + 0.018116334453225136, + 0.007874038070440292, + 0.0038095316849648952, + 0.0059037297032773495, + 0.005947572644799948, + 0.015752019360661507, + 0.008521863259375095, + 0.005063407123088837, + 0.031151888892054558, + 0.007509914226830006, + 0.004845531657338142, + 0.0070747858844697475, + 0.009412351995706558, + 0.027759257704019547, + 0.004906840156763792, + 0.0043407101184129715, + 0.0053394814021885395, + 0.0029096903745085, + 0.0038382578641176224, + 0.0055638886988162994, + 0.0035775001160800457, + 0.005136523395776749, + 0.01606890931725502, + 0.012466889806091785, + 0.019463758915662766, + 0.01559374574571848, + 0.018204400315880775, + 0.015127856284379959, + 0.014800531789660454, + 0.021115826442837715, + 0.008001252077519894, + 0.012842312455177307, + 0.009371579624712467, + 0.012581100687384605, + 0.011279413476586342, + 0.008985054679214954, + 0.008518679067492485, + 0.013672683387994766, + 0.004582660272717476, + 0.0033246083185076714, + 0.0044452110305428505, + 0.0033225733786821365, + 0.003201503772288561, + 0.008461209945380688, + 0.004278861451894045, + 0.0062446799129247665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.004093007650226355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured with the precision of ultra-high-speed photography, this extreme macro close-up slowly traces the richly textured, deep green surface of a rare spiny bamboo stalk. Other bamboo stems blur in the background, while soft golden morning light filters through the treetops, creating a dreamy, verdant bokeh effect and casting a delicate rim light across the scene. The camera's focus falls on a cluster of small, tightly sealed flower buds emerging from a spiny node on the stalk. In breathtaking, ultra-slow detail, one bud begins to tremble and swell, its outer skin splitting open in an elegant, explosive manner, revealing the intricate structure within. The delicate, paper-thin, cream-colored florets unfold in a fluid, almost liquid motion. Simultaneously, a cluster of thin, thread-like stamens burst forth, their pollen releasing a shimmering cloud of tiny golden pollen that hangs fleetingly in the still air. The camera meticulously tracks this rapid, graceful eruption, transforming the complex movements, normally invisible to the naked eye, into a spectacle of natural art, evoking a profound awe and the quiet, powerful beauty of a rare natural phenomenon.", + "seed": 15246, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_1/index_2484_seed_15246_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.005826318869367242, + "vlm_logprob": -0.005826318869367242, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.005826318869367242, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011460306122899055 + ], + "opposite_logprob_tokens": [ + -4.761460304260254 + ], + "logprob_reward": -0.011460306122899055, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011460306122899055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13505780696868896, + 0.07163002341985703, + 0.05338305979967117, + 0.08105496317148209, + 0.050343532115221024, + 0.032819945365190506, + 0.06202460825443268, + 0.10871972143650055, + 0.3316636383533478, + 0.07330270111560822 + ], + "credit_map": [ + 0.01918496936559677, + 0.020752018317580223, + 0.024492669850587845, + 0.01045011542737484, + 0.011263126507401466, + 0.010272243991494179, + 0.023342126980423927, + 0.015300541184842587, + 0.008793872781097889, + 0.010653317905962467, + 0.00731512950733304, + 0.0070721982046961784, + 0.008931687101721764, + 0.008610287681221962, + 0.0144574623554945, + 0.005796070210635662, + 0.008092673495411873, + 0.009294259361922741, + 0.006231344770640135, + 0.005138136446475983, + 0.0058870683424174786, + 0.003743943525478244, + 0.006987541448324919, + 0.00800809171050787, + 0.010763786733150482, + 0.012837053276598454, + 0.008688896894454956, + 0.008955108001828194, + 0.008945387788116932, + 0.0081002963706851, + 0.013798212632536888, + 0.008966227993369102, + 0.007477296981960535, + 0.008769679814577103, + 0.010274779051542282, + 0.006067320238798857, + 0.0038070690352469683, + 0.003702297108247876, + 0.005882060155272484, + 0.004363029729574919, + 0.0049239168874919415, + 0.004124196711927652, + 0.007030203007161617, + 0.0029814785812050104, + 0.002582319313660264, + 0.003383223433047533, + 0.0051672994159162045, + 0.002627308713272214, + 0.007682385388761759, + 0.011426744051277637, + 0.008576535619795322, + 0.004706311039626598, + 0.0067140148021280766, + 0.005149906035512686, + 0.008174285292625427, + 0.009594426490366459, + 0.010784342885017395, + 0.02430223673582077, + 0.015551978722214699, + 0.0142578249797225, + 0.009611282497644424, + 0.0087056839838624, + 0.015642711892724037, + 0.009863656014204025, + 0.04624054953455925, + 0.10731617361307144, + 0.07877643406391144, + 0.019423196092247963, + 0.023085199296474457, + 0.019319316372275352, + 0.015476412139832973, + 0.02202635630965233, + 0.009475630708038807, + 0.00982185173779726, + 0.011544049717485905, + 0.012833788059651852, + 0.006468629464507103, + 0.006350495852530003, + 0.007833367213606834, + 0.008974891155958176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.011460306122899055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003203262807801366 + ], + "opposite_logprob_tokens": [ + -5.753203392028809 + ], + "logprob_reward": -0.003203262807801366, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003203262807801366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25183501839637756, + 0.07617811113595963, + 0.040965691208839417, + 0.07365671545267105, + 0.06709393858909607, + 0.049921076744794846, + 0.03722037002444267, + 0.11614170670509338, + 0.19812968373298645, + 0.08885765820741653 + ], + "credit_map": [ + 0.02887645736336708, + 0.03531213849782944, + 0.03321769833564758, + 0.027799975126981735, + 0.03162654861807823, + 0.024848338216543198, + 0.036357440054416656, + 0.03379640728235245, + 0.008525359444320202, + 0.009870546869933605, + 0.008585356175899506, + 0.007951621897518635, + 0.010260112583637238, + 0.009784978814423084, + 0.014577126130461693, + 0.006623012945055962, + 0.0066667539067566395, + 0.0063780066557228565, + 0.005670114420354366, + 0.003523447085171938, + 0.003602325916290283, + 0.0034305984154343605, + 0.004576878156512976, + 0.007117568049579859, + 0.009289859794080257, + 0.012186810374259949, + 0.008652458898723125, + 0.009563250467181206, + 0.008111633360385895, + 0.007948200218379498, + 0.009519584476947784, + 0.008384917862713337, + 0.010079378262162209, + 0.010106856003403664, + 0.008784088306128979, + 0.0069185895845294, + 0.00923952180892229, + 0.005423606839030981, + 0.009631291031837463, + 0.006910607218742371, + 0.007563695777207613, + 0.00636797770857811, + 0.00821059849113226, + 0.0055753313936293125, + 0.005710603203624487, + 0.004270339384675026, + 0.007344694342464209, + 0.00487783458083868, + 0.00518031744286418, + 0.005074719432741404, + 0.005200695246458054, + 0.0033883797004818916, + 0.0029916532803326845, + 0.0032616518437862396, + 0.006449190434068441, + 0.00567376334220171, + 0.012954764999449253, + 0.023417441174387932, + 0.018380936235189438, + 0.014481853693723679, + 0.010118326172232628, + 0.01144773792475462, + 0.011446447111666203, + 0.013894197531044483, + 0.026947738602757454, + 0.061129938811063766, + 0.04670156538486481, + 0.013517884537577629, + 0.013590143993496895, + 0.012356665916740894, + 0.010426376946270466, + 0.013459382578730583, + 0.012042890302836895, + 0.011254457756876945, + 0.013918108306825161, + 0.017191939055919647, + 0.007901063188910484, + 0.006678135599941015, + 0.009591586887836456, + 0.010279474779963493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.003203262807801366 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022038952447474003 + ], + "opposite_logprob_tokens": [ + -6.127203941345215 + ], + "logprob_reward": -0.0022038952447474003, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022038952447474003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06399668008089066, + 0.062488626688718796, + 0.022558413445949554, + 0.06702527403831482, + 0.10306145250797272, + 0.028063492849469185, + 0.05580206215381622, + 0.09383837133646011, + 0.4013437330722809, + 0.10182192176580429 + ], + "credit_map": [ + 0.007790770847350359, + 0.012499953620135784, + 0.008062723092734814, + 0.006026708986610174, + 0.006626031827181578, + 0.005879017524421215, + 0.011377050541341305, + 0.0057344199158251286, + 0.010327189229428768, + 0.005914605688303709, + 0.01291936170309782, + 0.004097416531294584, + 0.003940870985388756, + 0.00701740151271224, + 0.007691363804042339, + 0.01058042049407959, + 0.0032725147902965546, + 0.004339591134339571, + 0.0026641404256224632, + 0.0021676456090062857, + 0.0015386904124170542, + 0.002135493094101548, + 0.0030145724304020405, + 0.0034257664810866117, + 0.009338676929473877, + 0.010007546283304691, + 0.008950786665081978, + 0.007433584425598383, + 0.007085776422172785, + 0.006812030915170908, + 0.009314228780567646, + 0.008082645013928413, + 0.020904770120978355, + 0.010112011805176735, + 0.029558436945080757, + 0.010187153704464436, + 0.00913216732442379, + 0.006237122230231762, + 0.010230268351733685, + 0.00669952342286706, + 0.0037161707878112793, + 0.0040936847217381, + 0.0032577316742390394, + 0.002930684946477413, + 0.003819886827841401, + 0.002517285058274865, + 0.00374567206017673, + 0.003982375841587782, + 0.006980538833886385, + 0.007875428535044193, + 0.010538612492382526, + 0.004283742979168892, + 0.006038805935531855, + 0.004319857805967331, + 0.007122574374079704, + 0.008642498403787613, + 0.009784283116459846, + 0.015622522681951523, + 0.011910228058695793, + 0.016148263588547707, + 0.008802455849945545, + 0.007859233766794205, + 0.013558442704379559, + 0.010152941569685936, + 0.052819281816482544, + 0.11959552764892578, + 0.09688441455364227, + 0.031253647059202194, + 0.026725677773356438, + 0.024729939177632332, + 0.020910413935780525, + 0.028424831107258797, + 0.011811480857431889, + 0.011455872096121311, + 0.01755913905799389, + 0.01843840442597866, + 0.0062992931343615055, + 0.01472277007997036, + 0.010707958601415157, + 0.010827002115547657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.0022038952447474003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630250882357359 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630250882357359, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630250882357359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05100273713469505, + 0.04432586207985878, + 0.01860361546278, + 0.031043678522109985, + 0.05259779095649719, + 0.024927891790866852, + 0.10420368611812592, + 0.5629082918167114, + 0.05399101600050926, + 0.05639544501900673 + ], + "credit_map": [ + 0.004833369981497526, + 0.010411609895527363, + 0.005492400377988815, + 0.0050354572013020515, + 0.008025637827813625, + 0.004889685194939375, + 0.007590293884277344, + 0.0047242785803973675, + 0.006941081956028938, + 0.00473850267007947, + 0.009250241331756115, + 0.0033252208959311247, + 0.003976025618612766, + 0.0046723089180886745, + 0.004958228208124638, + 0.006464253179728985, + 0.0030429577454924583, + 0.0027696203906089067, + 0.002530043013393879, + 0.0015956393908709288, + 0.001785270986147225, + 0.001584526733495295, + 0.0022628065198659897, + 0.0030327492859214544, + 0.006260077003389597, + 0.005032279063016176, + 0.005172534380108118, + 0.0026501405518501997, + 0.0031525571830570698, + 0.002332358853891492, + 0.002668675035238266, + 0.003775059012696147, + 0.010795417241752148, + 0.004394259303808212, + 0.01251908391714096, + 0.00502944178879261, + 0.005299041513353586, + 0.004318821243941784, + 0.0054445876739919186, + 0.004797139670699835, + 0.003541559912264347, + 0.004427476320415735, + 0.004017534665763378, + 0.0027021500281989574, + 0.0025873598642647266, + 0.0019393910188227892, + 0.002583606168627739, + 0.003128815209493041, + 0.012632561847567558, + 0.01859920658171177, + 0.015616449527442455, + 0.009348908439278603, + 0.010699078440666199, + 0.010166934691369534, + 0.00989541970193386, + 0.017245130613446236, + 0.07233936339616776, + 0.12217511981725693, + 0.07740192860364914, + 0.06209353730082512, + 0.04711565747857094, + 0.036364100873470306, + 0.07970339804887772, + 0.06571517139673233, + 0.010128230787813663, + 0.009024851955473423, + 0.0075933728367090225, + 0.007067735306918621, + 0.004242341499775648, + 0.0039407857693731785, + 0.005800257436931133, + 0.006193442270159721, + 0.007395876571536064, + 0.009981661103665829, + 0.007217667996883392, + 0.0101858489215374, + 0.005415394902229309, + 0.004471526015549898, + 0.004952524788677692, + 0.0067749470472335815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.007630250882357359 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633879289031029 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004633879289031029, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633879289031029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17035672068595886, + 0.17970426380634308, + 0.05296231806278229, + 0.08984920382499695, + 0.12538030743598938, + 0.05038604512810707, + 0.053376320749521255, + 0.0955180898308754, + 0.14363405108451843, + 0.03883266821503639 + ], + "credit_map": [ + 0.023868821561336517, + 0.03873700648546219, + 0.016488635912537575, + 0.014866752550005913, + 0.019963335245847702, + 0.013717742636799812, + 0.029251491650938988, + 0.013462936505675316, + 0.03174744173884392, + 0.026038160547614098, + 0.025304678827524185, + 0.01665329560637474, + 0.014390841126441956, + 0.015833038836717606, + 0.02258385345339775, + 0.027152948081493378, + 0.00873392540961504, + 0.008888859301805496, + 0.007401542272418737, + 0.004599894396960735, + 0.004960906691849232, + 0.0042787762358784676, + 0.005996620748192072, + 0.008101791143417358, + 0.01239018328487873, + 0.01416768878698349, + 0.00990300253033638, + 0.009903685189783573, + 0.00981230940669775, + 0.008548798970878124, + 0.014429987408220768, + 0.010693557560443878, + 0.020365126430988312, + 0.02119486965239048, + 0.018310846760869026, + 0.017395902425050735, + 0.012425869703292847, + 0.009614771232008934, + 0.013694385066628456, + 0.01237853430211544, + 0.00820391159504652, + 0.005880600772798061, + 0.008046240545809269, + 0.005505170673131943, + 0.0038086718413978815, + 0.005244819447398186, + 0.007485542446374893, + 0.006211091764271259, + 0.007192706689238548, + 0.009702463634312153, + 0.006595224142074585, + 0.005021913908421993, + 0.004517117515206337, + 0.003983610309660435, + 0.008350862190127373, + 0.008012425154447556, + 0.01390500646084547, + 0.016055764630436897, + 0.014332451857626438, + 0.00844412948936224, + 0.009387905709445477, + 0.0064840964041650295, + 0.0171914491802454, + 0.009717287495732307, + 0.02275099605321884, + 0.040626637637615204, + 0.02568039856851101, + 0.013088880106806755, + 0.009821728803217411, + 0.008293197490274906, + 0.009761177003383636, + 0.013611032627522945, + 0.0048026107251644135, + 0.004886073525995016, + 0.006111458875238895, + 0.006095740478485823, + 0.003613585839048028, + 0.0044981203973293304, + 0.00354914297349751, + 0.005275933537632227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.004633879289031029 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured with the precision of ultra-high-speed photography, this extreme macro close-up slowly traces the richly textured, deep green surface of a rare spiny bamboo stalk. Other bamboo stems blur in the background, while soft golden morning light filters through the treetops, creating a dreamy, verdant bokeh effect and casting a delicate rim light across the scene. The camera's focus falls on a cluster of small, tightly sealed flower buds emerging from a spiny node on the stalk. In breathtaking, ultra-slow detail, one bud begins to tremble and swell, its outer skin splitting open in an elegant, explosive manner, revealing the intricate structure within. The delicate, paper-thin, cream-colored florets unfold in a fluid, almost liquid motion. Simultaneously, a cluster of thin, thread-like stamens burst forth, their pollen releasing a shimmering cloud of tiny golden pollen that hangs fleetingly in the still air. The camera meticulously tracks this rapid, graceful eruption, transforming the complex movements, normally invisible to the naked eye, into a spectacle of natural art, evoking a profound awe and the quiet, powerful beauty of a rare natural phenomenon.", + "seed": 15247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_1/index_2484_seed_15247_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011589212901890278, + "vlm_logprob": -0.011589212901890278, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.011589212901890278, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012257367372512817 + ], + "opposite_logprob_tokens": [ + -4.7622575759887695 + ], + "logprob_reward": -0.012257367372512817, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012257367372512817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2422097623348236, + 0.10733325034379959, + 0.2034413367509842, + 0.0530574694275856, + 0.05692380294203758, + 0.05358172580599785, + 0.04996979981660843, + 0.05925491079688072, + 0.06266500800848007, + 0.11156295984983444 + ], + "credit_map": [ + 0.018819108605384827, + 0.05667104572057724, + 0.037866733968257904, + 0.022832335904240608, + 0.014513835310935974, + 0.020508481189608574, + 0.024155613034963608, + 0.04684258997440338, + 0.018802255392074585, + 0.015756580978631973, + 0.012253783643245697, + 0.013806149363517761, + 0.012744282372295856, + 0.008564830757677555, + 0.015523936599493027, + 0.009881420992314816, + 0.031083092093467712, + 0.024822769686579704, + 0.027695968747138977, + 0.01942897029221058, + 0.02584187127649784, + 0.01751025766134262, + 0.043852876871824265, + 0.013205506838858128, + 0.007799724116921425, + 0.00787136796861887, + 0.006042880471795797, + 0.0032045275438576937, + 0.004565997514873743, + 0.0038469419814646244, + 0.008194026537239552, + 0.01153200026601553, + 0.005774831399321556, + 0.007497059181332588, + 0.010316997766494751, + 0.005049009341746569, + 0.008123946376144886, + 0.005123375449329615, + 0.011006402783095837, + 0.00403217738494277, + 0.004217928741127253, + 0.008735690265893936, + 0.00859169103205204, + 0.006919104605913162, + 0.0052398089319467545, + 0.005444414913654327, + 0.008788777515292168, + 0.005644307937473059, + 0.005086714401841164, + 0.0057712881825864315, + 0.008125324733555317, + 0.0050775036215782166, + 0.005276859272271395, + 0.005716253072023392, + 0.009530486539006233, + 0.005385367665439844, + 0.0070482343435287476, + 0.008934525772929192, + 0.01163327507674694, + 0.004403922241181135, + 0.0052313790656626225, + 0.006348641123622656, + 0.009776055812835693, + 0.005878873635083437, + 0.00690237432718277, + 0.007371187675744295, + 0.008556576445698738, + 0.00814641360193491, + 0.005924331024289131, + 0.008130362257361412, + 0.011188962496817112, + 0.006444798316806555, + 0.013088938780128956, + 0.013399672694504261, + 0.013138863258063793, + 0.01045894157141447, + 0.012350158765912056, + 0.01135710347443819, + 0.025233197957277298, + 0.012536080554127693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.012257367372512817 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653285913169384 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.008653285913169384, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653285913169384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17002016305923462, + 0.08145720511674881, + 0.2617758810520172, + 0.06463492661714554, + 0.03363799676299095, + 0.023009466007351875, + 0.060972921550273895, + 0.02091122977435589, + 0.099319227039814, + 0.18426091969013214 + ], + "credit_map": [ + 0.017174581065773964, + 0.02450823038816452, + 0.041115764528512955, + 0.01825251802802086, + 0.015136445872485638, + 0.013472984544932842, + 0.020395729690790176, + 0.019963927567005157, + 0.006234065163880587, + 0.007331331260502338, + 0.01086299866437912, + 0.008847489021718502, + 0.012798728421330452, + 0.007967249490320683, + 0.02041596919298172, + 0.006999382749199867, + 0.03789593651890755, + 0.03906656429171562, + 0.02690136805176735, + 0.02845671772956848, + 0.03828122094273567, + 0.022756509482860565, + 0.052380457520484924, + 0.01603712886571884, + 0.007390111684799194, + 0.014883963391184807, + 0.006767641752958298, + 0.0030746806878596544, + 0.0044882153160870075, + 0.0031941316556185484, + 0.006394270341843367, + 0.01844191364943981, + 0.0032195360399782658, + 0.004014833364635706, + 0.0036940579302608967, + 0.0032402880024164915, + 0.005518704652786255, + 0.003770424285903573, + 0.006965493783354759, + 0.0032146605663001537, + 0.0020667335484176874, + 0.0030169833917170763, + 0.0034938848111778498, + 0.0019152763998135924, + 0.00305271171964705, + 0.0019846430514007807, + 0.004989558830857277, + 0.0024896764662116766, + 0.004267738666385412, + 0.00962271448224783, + 0.011297381483018398, + 0.005173835903406143, + 0.007714476902037859, + 0.005896799731999636, + 0.008606542833149433, + 0.008393436670303345, + 0.002953791758045554, + 0.001730632153339684, + 0.0027625630609691143, + 0.0016024805372580886, + 0.002866624388843775, + 0.0020657696295529604, + 0.00514275161549449, + 0.0017866176785901189, + 0.009814298711717129, + 0.009997701272368431, + 0.013995694927871227, + 0.008479530923068523, + 0.01252980250865221, + 0.017711808905005455, + 0.01870599389076233, + 0.008084399625658989, + 0.019681792706251144, + 0.022582916542887688, + 0.024875449016690254, + 0.017528289929032326, + 0.02183055318892002, + 0.019695797935128212, + 0.04136178269982338, + 0.016704346984624863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.008653285913169384 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209755696356297 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209755696356297, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209755696356297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16368940472602844, + 0.0801239162683487, + 0.08473529666662216, + 0.08441794663667679, + 0.09251248836517334, + 0.07104092091321945, + 0.12846499681472778, + 0.034295108169317245, + 0.0863901898264885, + 0.17432980239391327 + ], + "credit_map": [ + 0.013027388602495193, + 0.040372204035520554, + 0.025341369211673737, + 0.013887442648410797, + 0.011570790782570839, + 0.01244228146970272, + 0.014361490495502949, + 0.03268643096089363, + 0.0069492473267018795, + 0.008173291571438313, + 0.008935729041695595, + 0.007759959436953068, + 0.012203548103570938, + 0.00771692767739296, + 0.022298432886600494, + 0.006086770910769701, + 0.011251451447606087, + 0.011346595361828804, + 0.018173178657889366, + 0.006331379991024733, + 0.006488730665296316, + 0.007865420542657375, + 0.011930975131690502, + 0.011347557418048382, + 0.012559783644974232, + 0.01671532914042473, + 0.008581016212701797, + 0.004449420142918825, + 0.0065964143723249435, + 0.0053421263583004475, + 0.00748627632856369, + 0.02268756739795208, + 0.006920165382325649, + 0.011270903050899506, + 0.008593560196459293, + 0.010054880753159523, + 0.014467467553913593, + 0.009071891196072102, + 0.02499833144247532, + 0.0071352822706103325, + 0.006642794236540794, + 0.009613865055143833, + 0.010563199408352375, + 0.004752774722874165, + 0.00812600553035736, + 0.008288513869047165, + 0.017474038526415825, + 0.005579723510891199, + 0.009962407872080803, + 0.01665649004280567, + 0.019785696640610695, + 0.014094844460487366, + 0.02054191753268242, + 0.013847951777279377, + 0.017110170796513557, + 0.016465499997138977, + 0.0040604472160339355, + 0.0033981790766119957, + 0.005013508256524801, + 0.00275738094933331, + 0.0049177962355315685, + 0.0036999466829001904, + 0.007276713382452726, + 0.003171134740114212, + 0.009166951291263103, + 0.011372776702046394, + 0.009785912930965424, + 0.011809288524091244, + 0.011441627517342567, + 0.009689411148428917, + 0.0162638071924448, + 0.006860402412712574, + 0.017276393249630928, + 0.021281102672219276, + 0.023202579468488693, + 0.020930476486682892, + 0.0223015695810318, + 0.018303290009498596, + 0.03681046515703201, + 0.014223914593458176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.014209755696356297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189996480941772 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.018189996480941772, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189996480941772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08153524994850159, + 0.09911483526229858, + 0.0678446888923645, + 0.04689154028892517, + 0.085552878677845, + 0.06120994687080383, + 0.03739982098340988, + 0.06496353447437286, + 0.06068987399339676, + 0.3947976529598236 + ], + "credit_map": [ + 0.00692041777074337, + 0.01240367442369461, + 0.01380127388983965, + 0.008736411109566689, + 0.01022962387651205, + 0.006081505678594112, + 0.011286940425634384, + 0.012075405567884445, + 0.017456810921430588, + 0.014656237326562405, + 0.010480500757694244, + 0.007819218561053276, + 0.013953453861176968, + 0.006675595883280039, + 0.021911924704909325, + 0.006161087658256292, + 0.006156108807772398, + 0.016408182680606842, + 0.01356185507029295, + 0.007141867186874151, + 0.0068666948936879635, + 0.004718302749097347, + 0.007932517677545547, + 0.005059159826487303, + 0.005804209969937801, + 0.008009391836822033, + 0.0053233313374221325, + 0.003172483993694186, + 0.004242961294949055, + 0.003421633504331112, + 0.007543398067355156, + 0.009374132379889488, + 0.006831368897110224, + 0.011103492230176926, + 0.009352645836770535, + 0.009216846898198128, + 0.011734513565897942, + 0.0082318764179945, + 0.021997926756739616, + 0.007084212731570005, + 0.00472794845700264, + 0.005792765878140926, + 0.006910312920808792, + 0.004393336363136768, + 0.008124533109366894, + 0.007332481909543276, + 0.01888970285654068, + 0.005038866773247719, + 0.00441778777167201, + 0.00462179072201252, + 0.0052566551603376865, + 0.0037860132288187742, + 0.004756176378577948, + 0.003931189421564341, + 0.00629277341067791, + 0.004337434656918049, + 0.006851351819932461, + 0.00804722961038351, + 0.0075253830291330814, + 0.005306213162839413, + 0.007135396357625723, + 0.01033915113657713, + 0.013855752535164356, + 0.005903051234781742, + 0.00563633069396019, + 0.005262361839413643, + 0.0078128632158041, + 0.0058623007498681545, + 0.006210257299244404, + 0.008004057221114635, + 0.016491588205099106, + 0.0054101101122796535, + 0.0407024621963501, + 0.047761447727680206, + 0.0530584417283535, + 0.04626481235027313, + 0.04797494038939476, + 0.03651024028658867, + 0.09444856643676758, + 0.028076760470867157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 4 + ], + "selected_reward": -0.018189996480941772 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a deep green spiny bamboo stalk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focus on a cluster of small, tightly sealed flower buds emerging from a spiny node?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a cream-colored floret split open and unfold in a fluid, liquid-like motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do thin, thread-like stamens burst forth releasing a shimmering cloud of tiny golden pollen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of blurred bamboo stems with soft golden bokeh and rim lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635659046471119 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635659046471119, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635659046471119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13752050697803497, + 0.14953982830047607, + 0.06419076025485992, + 0.2565279006958008, + 0.044851139187812805, + 0.14901167154312134, + 0.06602204591035843, + 0.026410037651658058, + 0.055864185094833374, + 0.050061970949172974 + ], + "credit_map": [ + 0.012308791279792786, + 0.014839480631053448, + 0.026687053963541985, + 0.018987661227583885, + 0.014955955557525158, + 0.013582036830484867, + 0.02312321588397026, + 0.013036323711276054, + 0.03579211235046387, + 0.03317762538790703, + 0.017142850905656815, + 0.01659080758690834, + 0.010929069481790066, + 0.010970944538712502, + 0.013906547799706459, + 0.011029871180653572, + 0.008628488518297672, + 0.00846891663968563, + 0.011130787432193756, + 0.0065559097565710545, + 0.008410004898905754, + 0.006751292385160923, + 0.009310049004852772, + 0.004935312084853649, + 0.028450358659029007, + 0.06393935531377792, + 0.0331939272582531, + 0.012906801886856556, + 0.015043525025248528, + 0.021459082141518593, + 0.017644643783569336, + 0.06389021128416061, + 0.004225242882966995, + 0.0064199320040643215, + 0.006516358349472284, + 0.004196091089397669, + 0.005667136982083321, + 0.00450681708753109, + 0.0077896700240671635, + 0.005529891233891249, + 0.012706820853054523, + 0.021183840930461884, + 0.020135613158345222, + 0.023032523691654205, + 0.012959289364516735, + 0.02243204601109028, + 0.018958278000354767, + 0.017603250220417976, + 0.005641504656523466, + 0.008165517821907997, + 0.010780121199786663, + 0.006840623449534178, + 0.00932756345719099, + 0.006897870451211929, + 0.010168931446969509, + 0.008199911564588547, + 0.0029096752405166626, + 0.002837701002135873, + 0.0032714810222387314, + 0.0025165160186588764, + 0.0030470045749098063, + 0.003433516714721918, + 0.005803929641842842, + 0.0025902134366333485, + 0.006778372917324305, + 0.006329669151455164, + 0.006990744266659021, + 0.007148454897105694, + 0.006781283766031265, + 0.005260324105620384, + 0.010227142833173275, + 0.006348190829157829, + 0.00514153391122818, + 0.005508044734597206, + 0.005400271620601416, + 0.007864645682275295, + 0.005936082452535629, + 0.00523972325026989, + 0.008332747034728527, + 0.006638924591243267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.004635659046471119 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_10/index_tensor([276, 276])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_10/index_tensor([276, 276])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..03f55218c3fbd08c578e5130008f91581e2fa84e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_10/index_tensor([276, 276])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot, captured with the breathtaking clarity of high-frame-rate video, moves perfectly parallel to a majestic horse galloping across a vast, open plain at explosive speed. The horse, a majestic stallion with a sleek, near-black coat, epitomizes power; its long, flowing mane and wake trail behind it like silk banners in the wind. As it rumbles over the earth, every sinew and muscle group in its powerful shoulders, hips, and legs ripples visibly beneath its smooth skin. Each powerful stride is captured with surreal clarity, hooves striking the dry earth and launching in perfectly synchronized sequence, kicking up plumes of delicate golden dust that hang in the air. The entire scene is bathed in the warm, low-angle light of the afternoon sun, which creates a bright rim light around the horse's body, highlighting the texture of its coat and individual dust particles against a backdrop of sun-bleached golden grass. The atmosphere is one of uninhibited freedom and raw, breathtaking power, and the smooth, almost surreal quality of the high frame rate capture adds a graceful, balletic quality to the immense force on display.", + "seed": 30402, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_10/index_276_seed_30402_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.057697067596018316, + "vlm_logprob": -0.057697067596018316, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.057697067596018316, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1528930515050888 + ], + "opposite_logprob_tokens": [ + -2.40289306640625 + ], + "logprob_reward": -0.1528930515050888, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1528930515050888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08166373521089554, + 0.09351953864097595, + 0.11643296480178833, + 0.14794610440731049, + 0.048985399305820465, + 0.07770188897848129, + 0.029911626130342484, + 0.2890249788761139, + 0.08239880204200745, + 0.03241497650742531 + ], + "credit_map": [ + 0.013929629698395729, + 0.01519961841404438, + 0.012676901184022427, + 0.006488957908004522, + 0.009339980781078339, + 0.0076657626777887344, + 0.007492723874747753, + 0.008870155550539494, + 0.009405959397554398, + 0.033173371106386185, + 0.0077165584079921246, + 0.008878697641193867, + 0.010337499901652336, + 0.00460822181776166, + 0.005881900433450937, + 0.013517331331968307, + 0.006764417048543692, + 0.01925916038453579, + 0.003588884836062789, + 0.03136727586388588, + 0.0294774379581213, + 0.006051282398402691, + 0.006295100320130587, + 0.013629406690597534, + 0.013391958549618721, + 0.011289723217487335, + 0.016648022457957268, + 0.007650173269212246, + 0.010682888329029083, + 0.011516714468598366, + 0.04526522755622864, + 0.031501404941082, + 0.004633247386664152, + 0.011367624625563622, + 0.0032715729903429747, + 0.006922714877873659, + 0.00795181468129158, + 0.0045683481730520725, + 0.005047030746936798, + 0.005223043728619814, + 0.005704627837985754, + 0.008599123917520046, + 0.0042617893777787685, + 0.0038122895639389753, + 0.02340041846036911, + 0.018344821408391, + 0.005089364014565945, + 0.008489454165101051, + 0.0028916194569319487, + 0.0044559091329574585, + 0.0031229222659021616, + 0.004226644057780504, + 0.0031704213470220566, + 0.004549908451735973, + 0.0031178074423223734, + 0.004376394208520651, + 0.030455412343144417, + 0.0371374748647213, + 0.04032143950462341, + 0.02552674151957035, + 0.042245812714099884, + 0.03457815200090408, + 0.060273464769124985, + 0.018486471846699715, + 0.0037389909848570824, + 0.009980360046029091, + 0.0065967910923063755, + 0.006590510252863169, + 0.02524181827902794, + 0.014860762283205986, + 0.006919310428202152, + 0.008470263332128525, + 0.0029095085337758064, + 0.002965128282085061, + 0.005500371102243662, + 0.005409958306699991, + 0.004679775331169367, + 0.0037027259822934866, + 0.003898459253832698, + 0.003349050646647811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.1528930515050888 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07904716581106186 + ], + "opposite_logprob_tokens": [ + -2.579047203063965 + ], + "logprob_reward": -0.07904716581106186, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07904716581106186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21857315301895142, + 0.11191617697477341, + 0.07027412205934525, + 0.04070230945944786, + 0.0412948839366436, + 0.17555171251296997, + 0.03631458058953285, + 0.08677632361650467, + 0.19326508045196533, + 0.025331595912575722 + ], + "credit_map": [ + 0.03750406205654144, + 0.03922285512089729, + 0.0315127931535244, + 0.018035439774394035, + 0.02106167934834957, + 0.02148362435400486, + 0.021682627499103546, + 0.02807008847594261, + 0.010330247692763805, + 0.031298935413360596, + 0.009608631953597069, + 0.012842717580497265, + 0.013636114075779915, + 0.007888478226959705, + 0.009254900738596916, + 0.017056163400411606, + 0.009031210094690323, + 0.00979747623205185, + 0.0038900261279195547, + 0.009270194917917252, + 0.01124818716198206, + 0.007944528944790363, + 0.007847257889807224, + 0.011245246976613998, + 0.004315191879868507, + 0.005417879670858383, + 0.003199296072125435, + 0.0026243608444929123, + 0.004102710634469986, + 0.0037575976457446814, + 0.012743663042783737, + 0.004541609901934862, + 0.00553360627964139, + 0.0059502930380403996, + 0.003654900472611189, + 0.004326050169765949, + 0.005231923423707485, + 0.005782026797533035, + 0.004257310181856155, + 0.006558773573487997, + 0.017200008034706116, + 0.03552720323204994, + 0.01882719062268734, + 0.014169851318001747, + 0.0252517219632864, + 0.03132890537381172, + 0.014752429910004139, + 0.018494417890906334, + 0.0035257807467132807, + 0.003982683178037405, + 0.006478168070316315, + 0.0041311196982860565, + 0.0035041773226112127, + 0.0058089448139071465, + 0.0036198459565639496, + 0.0052638608030974865, + 0.007146521005779505, + 0.012133544310927391, + 0.010595782659947872, + 0.008318072184920311, + 0.015259718522429466, + 0.01215980015695095, + 0.014278166927397251, + 0.006884719245135784, + 0.013107772916555405, + 0.022709259763360023, + 0.02489255741238594, + 0.020296430215239525, + 0.046517565846443176, + 0.030491400510072708, + 0.01552564837038517, + 0.01972445473074913, + 0.0023201999720185995, + 0.002128869527950883, + 0.0038736823480576277, + 0.003624156815931201, + 0.005419491790235043, + 0.003232391085475683, + 0.0026516669895499945, + 0.002081140410155058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.07904716581106186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205760322511196 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.014205760322511196, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205760322511196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13078029453754425, + 0.06763654202222824, + 0.044852279126644135, + 0.19064250588417053, + 0.07850397378206253, + 0.10642363876104355, + 0.1895606368780136, + 0.0609971284866333, + 0.09218799322843552, + 0.0384150855243206 + ], + "credit_map": [ + 0.018602369353175163, + 0.024131227284669876, + 0.021091826260089874, + 0.014147431589663029, + 0.011850039474666119, + 0.016066793352365494, + 0.009731942787766457, + 0.01515864860266447, + 0.006119184661656618, + 0.01843791827559471, + 0.00694579491391778, + 0.006693733390420675, + 0.006933512631803751, + 0.006113153882324696, + 0.005896720569580793, + 0.010496513918042183, + 0.0037160939536988735, + 0.00976735632866621, + 0.002575305989012122, + 0.008245682343840599, + 0.00888624507933855, + 0.003652427112683654, + 0.003591001033782959, + 0.004418163560330868, + 0.012558188289403915, + 0.013316800817847252, + 0.016256257891654968, + 0.007329344283789396, + 0.01268319133669138, + 0.013087780214846134, + 0.06326519697904587, + 0.052145734429359436, + 0.008217503316700459, + 0.018313027918338776, + 0.005368353798985481, + 0.010191108100116253, + 0.012332032434642315, + 0.007105466444045305, + 0.007556972559541464, + 0.009419494308531284, + 0.012792747467756271, + 0.01203733216971159, + 0.015025983564555645, + 0.008606384508311749, + 0.018527401611208916, + 0.016429094597697258, + 0.009060211479663849, + 0.013944464735686779, + 0.02511834353208542, + 0.02774975448846817, + 0.021047018468379974, + 0.02110106498003006, + 0.01899810880422592, + 0.0315411277115345, + 0.01606496050953865, + 0.027940237894654274, + 0.005598981864750385, + 0.007700488902628422, + 0.009356174618005753, + 0.00412026047706604, + 0.010212165303528309, + 0.008527769707143307, + 0.010987062938511372, + 0.004494213964790106, + 0.00436736736446619, + 0.01248292624950409, + 0.007082595024257898, + 0.007776150479912758, + 0.028583087027072906, + 0.01666223257780075, + 0.007275666110217571, + 0.007957957684993744, + 0.003463089233264327, + 0.004299360327422619, + 0.0033778012730181217, + 0.004099954850971699, + 0.011175998486578465, + 0.006224959157407284, + 0.0027072906959801912, + 0.003066628472879529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.014205760322511196 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803065583109856 + ], + "opposite_logprob_tokens": [ + -3.5298030376434326 + ], + "logprob_reward": -0.029803065583109856, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803065583109856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14424601197242737, + 0.1318386346101761, + 0.047161128371953964, + 0.032375138252973557, + 0.03257325664162636, + 0.2896343171596527, + 0.01586996763944626, + 0.22948850691318512, + 0.03895081207156181, + 0.037862226366996765 + ], + "credit_map": [ + 0.022380346432328224, + 0.029033439233899117, + 0.022990357130765915, + 0.015334272757172585, + 0.010739319957792759, + 0.014773714356124401, + 0.00969764869660139, + 0.01929691806435585, + 0.01364960614591837, + 0.0424920953810215, + 0.01037271972745657, + 0.01362844929099083, + 0.014156773686408997, + 0.0097390441223979, + 0.009168364107608795, + 0.018631575629115105, + 0.002905890578404069, + 0.01140605565160513, + 0.0018013898516073823, + 0.01687062531709671, + 0.0039061098359525204, + 0.002536497777327895, + 0.0033426168374717236, + 0.004391942173242569, + 0.003466270864009857, + 0.004813431762158871, + 0.0039344667457044125, + 0.0027388546150177717, + 0.0026689462829381227, + 0.0029415700118988752, + 0.004928296431899071, + 0.006883301306515932, + 0.0042669023387134075, + 0.007033660542219877, + 0.002162016462534666, + 0.003992273472249508, + 0.003632578067481518, + 0.0023107435554265976, + 0.0038081032689660788, + 0.005366981960833073, + 0.045420143753290176, + 0.02947060391306877, + 0.016700005158782005, + 0.010802112519741058, + 0.051074687391519547, + 0.08688389509916306, + 0.01635633036494255, + 0.0329265259206295, + 0.0017683171899989247, + 0.002653119619935751, + 0.0016678815009072423, + 0.0019681809935718775, + 0.002293632598593831, + 0.0019365486223250628, + 0.0014180076541379094, + 0.0021642807405442, + 0.023477166891098022, + 0.025865918025374413, + 0.027578461915254593, + 0.021860092878341675, + 0.0389530248939991, + 0.02753308042883873, + 0.046351108700037, + 0.01786963827908039, + 0.004444008693099022, + 0.005000252276659012, + 0.004598569590598345, + 0.0025288371834903955, + 0.008655720390379429, + 0.006007731426507235, + 0.0037835168186575174, + 0.0039321742951869965, + 0.0031386108603328466, + 0.0038850728888064623, + 0.005352906882762909, + 0.005863289348781109, + 0.005799105856567621, + 0.004779853392392397, + 0.0042524319142103195, + 0.004790955688804388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.029803065583109856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536294758319855 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536294758319855, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536294758319855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17184478044509888, + 0.08109335601329803, + 0.09683608263731003, + 0.1799997240304947, + 0.08959127217531204, + 0.1630110740661621, + 0.02392827905714512, + 0.06296616047620773, + 0.07420752197504044, + 0.0565216988325119 + ], + "credit_map": [ + 0.024622002616524696, + 0.023717273026704788, + 0.032705046236515045, + 0.014834906905889511, + 0.01645655184984207, + 0.014105417765676975, + 0.0251017976552248, + 0.02030179835855961, + 0.007956961169838905, + 0.021164607256650925, + 0.00829792395234108, + 0.011507144197821617, + 0.008651351556181908, + 0.005446466151624918, + 0.008439852856099606, + 0.009629053995013237, + 0.005760468542575836, + 0.011882614344358444, + 0.004574343096464872, + 0.03377249836921692, + 0.016027159988880157, + 0.006160041783004999, + 0.006748444866389036, + 0.011910512112081051, + 0.011481448076665401, + 0.009164918214082718, + 0.008702595718204975, + 0.006197627633810043, + 0.010725210420787334, + 0.01343197375535965, + 0.06763236224651337, + 0.052663594484329224, + 0.009842502884566784, + 0.018697330728173256, + 0.005567448679357767, + 0.011392312124371529, + 0.013702034950256348, + 0.008345786482095718, + 0.008114883676171303, + 0.013928975909948349, + 0.033015426248311996, + 0.024218104779720306, + 0.011985400691628456, + 0.009712055325508118, + 0.012758755125105381, + 0.04323991388082504, + 0.008860672824084759, + 0.019220765680074692, + 0.002908777678385377, + 0.003686669282615185, + 0.0026905126869678497, + 0.0029600702691823244, + 0.0026619313284754753, + 0.002453140215948224, + 0.0022052477579563856, + 0.0043619307689368725, + 0.005197568330913782, + 0.0095827616751194, + 0.00564188975840807, + 0.006046472582966089, + 0.009910963475704193, + 0.008712703362107277, + 0.012969705276191235, + 0.004904099274426699, + 0.004984429571777582, + 0.010401055216789246, + 0.006126525811851025, + 0.00891194399446249, + 0.01835743524134159, + 0.012243414297699928, + 0.005266265477985144, + 0.007916459813714027, + 0.004996000323444605, + 0.007208812981843948, + 0.004461152013391256, + 0.006703807506710291, + 0.014769257046282291, + 0.008600120432674885, + 0.004065132234245539, + 0.005717417225241661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.012536294758319855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, captured with the breathtaking clarity of high-frame-rate video, moves perfectly parallel to a majestic horse galloping across a vast, open plain at explosive speed. The horse, a majestic stallion with a sleek, near-black coat, epitomizes power; its long, flowing mane and wake trail behind it like silk banners in the wind. As it rumbles over the earth, every sinew and muscle group in its powerful shoulders, hips, and legs ripples visibly beneath its smooth skin. Each powerful stride is captured with surreal clarity, hooves striking the dry earth and launching in perfectly synchronized sequence, kicking up plumes of delicate golden dust that hang in the air. The entire scene is bathed in the warm, low-angle light of the afternoon sun, which creates a bright rim light around the horse's body, highlighting the texture of its coat and individual dust particles against a backdrop of sun-bleached golden grass. The atmosphere is one of uninhibited freedom and raw, breathtaking power, and the smooth, almost surreal quality of the high frame rate capture adds a graceful, balletic quality to the immense force on display.", + "seed": 30403, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_10/index_276_seed_30403_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05456401184201241, + "vlm_logprob": -0.05456401184201241, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.05456401184201241, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17504997551441193 + ], + "opposite_logprob_tokens": [ + -2.4250500202178955 + ], + "logprob_reward": -0.17504997551441193, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.17504997551441193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18552887439727783, + 0.0443728007376194, + 0.030041389167308807, + 0.029984742403030396, + 0.07800932228565216, + 0.41712361574172974, + 0.09146393835544586, + 0.0249992273747921, + 0.07940429449081421, + 0.01907176338136196 + ], + "credit_map": [ + 0.045238882303237915, + 0.013119145296514034, + 0.05663701519370079, + 0.012195359915494919, + 0.014344319701194763, + 0.014855182729661465, + 0.014136336743831635, + 0.015002654865384102, + 0.007262903731316328, + 0.003998198546469212, + 0.009033434092998505, + 0.00426053861156106, + 0.007021575700491667, + 0.004218438174575567, + 0.004218804184347391, + 0.004358911421149969, + 0.005121495574712753, + 0.0020911432802677155, + 0.007856262847781181, + 0.00233894563280046, + 0.0048938654363155365, + 0.002900946419686079, + 0.0026939117815345526, + 0.002144818427041173, + 0.005546923726797104, + 0.0030819247476756573, + 0.0065102120861411095, + 0.0026719304732978344, + 0.004834888037294149, + 0.002880566753447056, + 0.0018878700211644173, + 0.0025704270228743553, + 0.011776451021432877, + 0.005195891950279474, + 0.009126354940235615, + 0.006834309548139572, + 0.006815334782004356, + 0.0049505229108035564, + 0.012756431475281715, + 0.02055402658879757, + 0.09188729524612427, + 0.17497146129608154, + 0.014183214865624905, + 0.018106523901224136, + 0.07551415264606476, + 0.011613413691520691, + 0.008706570602953434, + 0.022141000255942345, + 0.01940828002989292, + 0.007876968942582607, + 0.01179595198482275, + 0.006918734405189753, + 0.010021017864346504, + 0.010142315179109573, + 0.012236405164003372, + 0.013064264319837093, + 0.0037573417648673058, + 0.0024901800788939, + 0.0050641982816159725, + 0.002572078024968505, + 0.002257233951240778, + 0.003215301549062133, + 0.003556478302925825, + 0.00208641798235476, + 0.014266019687056541, + 0.008219169452786446, + 0.00863711629062891, + 0.00597048457711935, + 0.009640734642744064, + 0.01651906967163086, + 0.006787245627492666, + 0.009364461526274681, + 0.0016011596890166402, + 0.003205098444595933, + 0.0023769475519657135, + 0.0038450693245977163, + 0.001982534769922495, + 0.0018998810555785894, + 0.0018832067726179957, + 0.002277868567034602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.17504997551441193 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03380666673183441 + ], + "opposite_logprob_tokens": [ + -3.408806562423706 + ], + "logprob_reward": -0.03380666673183441, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03380666673183441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0644611269235611, + 0.06519017368555069, + 0.06458767503499985, + 0.10122551023960114, + 0.19885008037090302, + 0.10033587366342545, + 0.07180211693048477, + 0.16406339406967163, + 0.14842791855335236, + 0.021056095138192177 + ], + "credit_map": [ + 0.009451167657971382, + 0.0075083449482917786, + 0.013786010444164276, + 0.0048765214160084724, + 0.006443880032747984, + 0.00770048052072525, + 0.0073463707230985165, + 0.0073483469896018505, + 0.0073464722372591496, + 0.007591818924993277, + 0.007673182524740696, + 0.006421537604182959, + 0.015497740358114243, + 0.010510362684726715, + 0.0056359050795435905, + 0.004513160325586796, + 0.014873214066028595, + 0.005821186117827892, + 0.01715286448597908, + 0.004779561422765255, + 0.006939914543181658, + 0.005430324003100395, + 0.0059621636755764484, + 0.003628440899774432, + 0.014223449863493443, + 0.013239171355962753, + 0.015821581706404686, + 0.010497733019292355, + 0.018134668469429016, + 0.011318037286400795, + 0.0073071992956101894, + 0.010683675296604633, + 0.02679865062236786, + 0.015793878585100174, + 0.020999850705266, + 0.014510015025734901, + 0.03830933943390846, + 0.011820102110505104, + 0.027308916673064232, + 0.043309323489665985, + 0.01680869795382023, + 0.022418705746531487, + 0.030869703739881516, + 0.007418207358568907, + 0.010794012807309628, + 0.0027790458407253027, + 0.004579230677336454, + 0.004668269772082567, + 0.01436837762594223, + 0.006473443005234003, + 0.006440944969654083, + 0.004013141617178917, + 0.01096805278211832, + 0.00722299050539732, + 0.006004082038998604, + 0.01631108671426773, + 0.025541435927152634, + 0.015330785885453224, + 0.015025620348751545, + 0.01393941231071949, + 0.041520193219184875, + 0.02132616937160492, + 0.016607867553830147, + 0.014771910384297371, + 0.027028700336813927, + 0.014104165136814117, + 0.016232021152973175, + 0.011555276811122894, + 0.017780520021915436, + 0.030188757926225662, + 0.01401921920478344, + 0.017519250512123108, + 0.0019496509339660406, + 0.0038386895321309566, + 0.002468772465363145, + 0.003259200369939208, + 0.0025991625152528286, + 0.001919918111525476, + 0.001955876825377345, + 0.0030648251995444298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.03380666673183441 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009793803095817566 + ], + "opposite_logprob_tokens": [ + -4.634793758392334 + ], + "logprob_reward": -0.009793803095817566, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009793803095817566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11386718600988388, + 0.06486150622367859, + 0.07844137400388718, + 0.052509959787130356, + 0.038199592381715775, + 0.16846410930156708, + 0.20022037625312805, + 0.10499854385852814, + 0.04982195049524307, + 0.1286153942346573 + ], + "credit_map": [ + 0.017352085560560226, + 0.01640053279697895, + 0.01784203015267849, + 0.006445314735174179, + 0.02526719681918621, + 0.011458802036941051, + 0.010208064690232277, + 0.008893162943422794, + 0.010639892891049385, + 0.005975144449621439, + 0.010014712810516357, + 0.00464619230479002, + 0.01638021692633629, + 0.006959781516343355, + 0.006069500464946032, + 0.0041760667227208614, + 0.01176634430885315, + 0.009922653436660767, + 0.015226360410451889, + 0.004674494732171297, + 0.015608743764460087, + 0.007438913453370333, + 0.008647586219012737, + 0.005156281404197216, + 0.007806679233908653, + 0.007699035108089447, + 0.007926294580101967, + 0.004304337315261364, + 0.011034118942916393, + 0.00556974159553647, + 0.0040619936771690845, + 0.004107760731130838, + 0.005735871847718954, + 0.002935558557510376, + 0.005788281559944153, + 0.00254521775059402, + 0.008499701507389545, + 0.003860427998006344, + 0.005530186463147402, + 0.003304349957033992, + 0.027221685275435448, + 0.03919203206896782, + 0.054915063083171844, + 0.011889559216797352, + 0.017466925084590912, + 0.003356565022841096, + 0.006972481496632099, + 0.0074498001486063, + 0.04533661529421806, + 0.01300410833209753, + 0.014236243441700935, + 0.012506532482802868, + 0.018576640635728836, + 0.016222506761550903, + 0.01718869060277939, + 0.06314902752637863, + 0.01987754926085472, + 0.01680331490933895, + 0.010750784538686275, + 0.006153874099254608, + 0.0196287352591753, + 0.013171257451176643, + 0.012158236466348171, + 0.006454783026129007, + 0.007899900898337364, + 0.007496723905205727, + 0.0058281319215893745, + 0.004016129765659571, + 0.006374828051775694, + 0.008679159916937351, + 0.004036156926304102, + 0.005490920040756464, + 0.012813110835850239, + 0.0284393560141325, + 0.014131884090602398, + 0.014626648277044296, + 0.017385201528668404, + 0.01181882806122303, + 0.011066315695643425, + 0.018334051594138145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.009793803095817566 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810453042387962 + ], + "opposite_logprob_tokens": [ + -3.28810453414917 + ], + "logprob_reward": -0.03810453042387962, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810453042387962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3283451795578003, + 0.04385451227426529, + 0.07791014760732651, + 0.04817114770412445, + 0.04042050614953041, + 0.1973164677619934, + 0.03961010277271271, + 0.04829450324177742, + 0.15402568876743317, + 0.02205175906419754 + ], + "credit_map": [ + 0.08436299860477448, + 0.026746118441224098, + 0.07664994895458221, + 0.023600926622748375, + 0.02837868593633175, + 0.0284899789839983, + 0.03078778274357319, + 0.029328761622309685, + 0.009006453678011894, + 0.0027240924537181854, + 0.013241690583527088, + 0.003323455573990941, + 0.0068349046632647514, + 0.002479840535670519, + 0.0033652731217443943, + 0.0028788095805794, + 0.015291711315512657, + 0.0065638297237455845, + 0.02294362150132656, + 0.004148541484028101, + 0.010880711488425732, + 0.008433816954493523, + 0.005306228995323181, + 0.0043416996486485004, + 0.006993505638092756, + 0.005863490514457226, + 0.007311214692890644, + 0.0050950865261256695, + 0.008808561600744724, + 0.0049614496529102325, + 0.0036048514302819967, + 0.005532992538064718, + 0.005546226631850004, + 0.004923448897898197, + 0.004445650614798069, + 0.00217045983299613, + 0.013598657213151455, + 0.0036083501763641834, + 0.0030873629730194807, + 0.0030403530690819025, + 0.05490479618310928, + 0.06729532033205032, + 0.02214479260146618, + 0.0071423668414354324, + 0.028327949345111847, + 0.004807681310921907, + 0.004211746156215668, + 0.008481825701892376, + 0.010258480906486511, + 0.0026504395063966513, + 0.003503277199342847, + 0.0017347324173897505, + 0.003304698271676898, + 0.0027508079074323177, + 0.0029088810551911592, + 0.01249879039824009, + 0.006075141951441765, + 0.006876942701637745, + 0.007046399172395468, + 0.004146161489188671, + 0.0049853515811264515, + 0.007020719349384308, + 0.006811232306063175, + 0.005332559812813997, + 0.027324656024575233, + 0.01567172072827816, + 0.017436129972338676, + 0.012579395435750484, + 0.017902424558997154, + 0.031316474080085754, + 0.013929934240877628, + 0.01786496676504612, + 0.0016825238708406687, + 0.003534723771736026, + 0.0019340034341439605, + 0.004499533679336309, + 0.002731120213866234, + 0.0020440397784113884, + 0.0020770267583429813, + 0.0035487895365804434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.03810453042387962 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160650834441185 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.0160650834441185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160650834441185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14939497411251068, + 0.06232348084449768, + 0.03128800541162491, + 0.029083779081702232, + 0.06247006356716156, + 0.4744492471218109, + 0.03022323176264763, + 0.029370112344622612, + 0.05776800960302353, + 0.07362907379865646 + ], + "credit_map": [ + 0.025216395035386086, + 0.015640627592802048, + 0.03182798624038696, + 0.011985541321337223, + 0.015361646190285683, + 0.016836192458868027, + 0.017652923241257668, + 0.014873656444251537, + 0.008617870509624481, + 0.006902088411152363, + 0.0103480014950037, + 0.0051177130080759525, + 0.013547730632126331, + 0.007630875334143639, + 0.005847988184541464, + 0.004311209078878164, + 0.005823301151394844, + 0.0021496331319212914, + 0.008201653137803078, + 0.00204264372587204, + 0.004830015357583761, + 0.0026758937165141106, + 0.0028943251818418503, + 0.002670540940016508, + 0.004892419092357159, + 0.0037741067353636026, + 0.005900116637349129, + 0.0025357117410749197, + 0.005088845733553171, + 0.00271377875469625, + 0.001937534543685615, + 0.00224126735702157, + 0.009594385512173176, + 0.004775465000420809, + 0.0076670218259096146, + 0.005452340468764305, + 0.005882927216589451, + 0.00414956035092473, + 0.007976501248776913, + 0.01697186380624771, + 0.09503406286239624, + 0.19915710389614105, + 0.023987984284758568, + 0.021459326148033142, + 0.08613982796669006, + 0.01338073331862688, + 0.009948235005140305, + 0.025341976433992386, + 0.00439531821757555, + 0.002538248198106885, + 0.0035989568568766117, + 0.001964719034731388, + 0.003801013110205531, + 0.004492256324738264, + 0.0042262147180736065, + 0.005206502508372068, + 0.004940203856676817, + 0.004175120964646339, + 0.002981516532599926, + 0.0025026430375874043, + 0.005710464436560869, + 0.0032096654176712036, + 0.0035998905077576637, + 0.0022506071254611015, + 0.008882389403879642, + 0.00713260006159544, + 0.006788141559809446, + 0.005043329205363989, + 0.00759646762162447, + 0.010914317332208157, + 0.00509627116844058, + 0.0063144913874566555, + 0.006842158269137144, + 0.010055741295218468, + 0.0064558363519608974, + 0.015170224010944366, + 0.007155917584896088, + 0.006784160155802965, + 0.006716371979564428, + 0.014448666013777256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.0160650834441185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, captured with the breathtaking clarity of high-frame-rate video, moves perfectly parallel to a majestic horse galloping across a vast, open plain at explosive speed. The horse, a majestic stallion with a sleek, near-black coat, epitomizes power; its long, flowing mane and wake trail behind it like silk banners in the wind. As it rumbles over the earth, every sinew and muscle group in its powerful shoulders, hips, and legs ripples visibly beneath its smooth skin. Each powerful stride is captured with surreal clarity, hooves striking the dry earth and launching in perfectly synchronized sequence, kicking up plumes of delicate golden dust that hang in the air. The entire scene is bathed in the warm, low-angle light of the afternoon sun, which creates a bright rim light around the horse's body, highlighting the texture of its coat and individual dust particles against a backdrop of sun-bleached golden grass. The atmosphere is one of uninhibited freedom and raw, breathtaking power, and the smooth, almost surreal quality of the high frame rate capture adds a graceful, balletic quality to the immense force on display.", + "seed": 30404, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_10/index_276_seed_30404_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.031834783405065535, + "vlm_logprob": -0.031834783405065535, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.031834783405065535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08593250066041946 + ], + "opposite_logprob_tokens": [ + -3.335932493209839 + ], + "logprob_reward": -0.08593250066041946, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.08593250066041946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4202916920185089, + 0.03292442485690117, + 0.045544661581516266, + 0.015632199123501778, + 0.12005792558193207, + 0.03770643100142479, + 0.09354594349861145, + 0.0956040546298027, + 0.020779477432370186, + 0.11791317164897919 + ], + "credit_map": [ + 0.038318756967782974, + 0.04166938364505768, + 0.012765023857355118, + 0.028056621551513672, + 0.09765186160802841, + 0.05893363431096077, + 0.0180317722260952, + 0.1248646229505539, + 0.0019809124059975147, + 0.006932049058377743, + 0.0030467903707176447, + 0.001769806258380413, + 0.00410249549895525, + 0.005318907089531422, + 0.0021297624334692955, + 0.007643702439963818, + 0.0029314891435205936, + 0.004061358980834484, + 0.0031357307452708483, + 0.01563369855284691, + 0.007928015664219856, + 0.004332434386014938, + 0.0031360043212771416, + 0.004385932348668575, + 0.001966501120477915, + 0.00194109370931983, + 0.00208804989233613, + 0.0010976361809298396, + 0.0022182874381542206, + 0.0022545147221535444, + 0.0026886577252298594, + 0.0013774598482996225, + 0.014143825508654118, + 0.006160943768918514, + 0.03720420226454735, + 0.009401111863553524, + 0.008753074333071709, + 0.03111102618277073, + 0.007443720940500498, + 0.005840024445205927, + 0.004661925602704287, + 0.002375624142587185, + 0.008837943896651268, + 0.0030844826251268387, + 0.0041611045598983765, + 0.006132254842668772, + 0.006105946842581034, + 0.0023471475578844547, + 0.016777928918600082, + 0.009422020986676216, + 0.014958621002733707, + 0.005762717220932245, + 0.01836150698363781, + 0.015262587927281857, + 0.006434308830648661, + 0.006566250696778297, + 0.0062807113863527775, + 0.006218889728188515, + 0.02733803167939186, + 0.008353408426046371, + 0.012907866388559341, + 0.006202156189829111, + 0.022470248863101006, + 0.005832738243043423, + 0.001367082353681326, + 0.0031200991943478584, + 0.0025007345248013735, + 0.0025363254826515913, + 0.0030175806023180485, + 0.002059696475043893, + 0.0033364377450197935, + 0.002841521752998233, + 0.01049950160086155, + 0.01596335880458355, + 0.017685841768980026, + 0.010327698662877083, + 0.021560903638601303, + 0.01762378215789795, + 0.014287898316979408, + 0.009964179247617722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.08593250066041946 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03376656025648117 + ], + "opposite_logprob_tokens": [ + -3.408766508102417 + ], + "logprob_reward": -0.03376656025648117, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03376656025648117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6492434740066528, + 0.05272689834237099, + 0.054822888225317, + 0.014478323049843311, + 0.05400308966636658, + 0.05459031090140343, + 0.03906017914414406, + 0.014874167740345001, + 0.013919525779783726, + 0.052281081676483154 + ], + "credit_map": [ + 0.06043795868754387, + 0.06317741423845291, + 0.0174116063863039, + 0.04330034926533699, + 0.1559276431798935, + 0.08980392664670944, + 0.02781674824655056, + 0.19136784970760345, + 0.0036624313797801733, + 0.009175249375402927, + 0.006033308804035187, + 0.0035131056793034077, + 0.00652677658945322, + 0.00834064744412899, + 0.004308381117880344, + 0.011167001910507679, + 0.0024013984948396683, + 0.00394927803426981, + 0.0035969337914139032, + 0.020287515595555305, + 0.010659638792276382, + 0.005398264154791832, + 0.002974910894408822, + 0.005554954055696726, + 0.002015097066760063, + 0.00199592811986804, + 0.002436554292216897, + 0.0011666648788377643, + 0.0019473384600132704, + 0.0015694407047703862, + 0.0022833431139588356, + 0.0010639589745551348, + 0.006520669441670179, + 0.003311421489343047, + 0.012566273100674152, + 0.004508263431489468, + 0.006969866808503866, + 0.012760219164192677, + 0.0039000273682177067, + 0.0034663493279367685, + 0.01352282427251339, + 0.003318269271403551, + 0.009851144626736641, + 0.0025013729464262724, + 0.006652899086475372, + 0.011976256035268307, + 0.003948351833969355, + 0.002819197019562125, + 0.005081906449049711, + 0.00375014403834939, + 0.005599744617938995, + 0.003269251435995102, + 0.009318836964666843, + 0.005004181992262602, + 0.003366585588082671, + 0.0036695320159196854, + 0.001548389787785709, + 0.0010365110356360674, + 0.002092314651235938, + 0.0021505204495042562, + 0.003385307500138879, + 0.0013988739810883999, + 0.0020592771470546722, + 0.001202974235638976, + 0.0010246599558740854, + 0.001897816895507276, + 0.00199471996165812, + 0.001301212003454566, + 0.0022094743326306343, + 0.001159242819994688, + 0.0020659419242292643, + 0.0022664591670036316, + 0.006655591074377298, + 0.009433960542082787, + 0.008441139943897724, + 0.0046877930872142315, + 0.008410424925386906, + 0.004225533455610275, + 0.006820335052907467, + 0.0036063087172806263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.03376656025648117 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205290004611015 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.014205290004611015, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205290004611015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41560983657836914, + 0.03390464931726456, + 0.11376290023326874, + 0.06281540542840958, + 0.05888393148779869, + 0.06559569388628006, + 0.06312514841556549, + 0.03628867492079735, + 0.032208722084760666, + 0.11780508607625961 + ], + "credit_map": [ + 0.03635018691420555, + 0.044545188546180725, + 0.012281102128326893, + 0.028423817828297615, + 0.09023800492286682, + 0.05810707062482834, + 0.020674141123890877, + 0.12499032914638519, + 0.00373008381575346, + 0.0059656105004251, + 0.003559735370799899, + 0.003674902953207493, + 0.004407144151628017, + 0.0037703546695411205, + 0.003184173023328185, + 0.005612647160887718, + 0.004978321027010679, + 0.006987770088016987, + 0.007234606426209211, + 0.043234482407569885, + 0.0232453141361475, + 0.009988551028072834, + 0.006351312156766653, + 0.011742541566491127, + 0.0061055379919707775, + 0.008826467208564281, + 0.009415945969522, + 0.004013030789792538, + 0.011898435652256012, + 0.007769228424876928, + 0.010378312319517136, + 0.004408449865877628, + 0.005035309121012688, + 0.00637049088254571, + 0.005117629189044237, + 0.00529967388138175, + 0.018808487802743912, + 0.008413727395236492, + 0.004447437822818756, + 0.005391179118305445, + 0.006519784685224295, + 0.007820520550012589, + 0.012160591781139374, + 0.004163980949670076, + 0.017304718494415283, + 0.007903059013187885, + 0.0061800358816981316, + 0.0035429969429969788, + 0.011699666269123554, + 0.004804583732038736, + 0.01014930009841919, + 0.003974313382059336, + 0.010276632383465767, + 0.011269044131040573, + 0.005031288601458073, + 0.005920319817960262, + 0.0037503333296626806, + 0.004177656024694443, + 0.004731121007353067, + 0.004308979492634535, + 0.006127943750470877, + 0.00266440212726593, + 0.0070481207221746445, + 0.003480119863525033, + 0.0019245569128543139, + 0.004184724297374487, + 0.007906388491392136, + 0.0034493908751755953, + 0.00493214838206768, + 0.0026134545914828777, + 0.00488963071256876, + 0.0023084271233528852, + 0.013245291076600552, + 0.013889767229557037, + 0.02003306709229946, + 0.01292917039245367, + 0.017452148720622063, + 0.011834551580250263, + 0.015657223761081696, + 0.012763863429427147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.014205290004611015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419647503644228 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.01419647503644228, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419647503644228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0771433413028717, + 0.13153544068336487, + 0.053350258618593216, + 0.03632471710443497, + 0.2823244333267212, + 0.07577095180749893, + 0.0860643982887268, + 0.028437405824661255, + 0.03871523216366768, + 0.19033388793468475 + ], + "credit_map": [ + 0.007004838436841965, + 0.014587976969778538, + 0.006381663028150797, + 0.007815733551979065, + 0.011705003678798676, + 0.009782749228179455, + 0.00807020254433155, + 0.011795162223279476, + 0.009613131172955036, + 0.026635557413101196, + 0.013784109614789486, + 0.008712626062333584, + 0.01752627268433571, + 0.01843506097793579, + 0.011583816260099411, + 0.02524486370384693, + 0.0033186820801347494, + 0.006526106968522072, + 0.004040414933115244, + 0.014589371159672737, + 0.008724641986191273, + 0.005069422069936991, + 0.0048758196644485, + 0.006205795798450708, + 0.004066260531544685, + 0.003566778963431716, + 0.006858091801404953, + 0.0024433473590761423, + 0.004242779221385717, + 0.006444901693612337, + 0.0054572587832808495, + 0.0032452938612550497, + 0.03161093592643738, + 0.01951673999428749, + 0.07046599686145782, + 0.02656627632677555, + 0.026910075917840004, + 0.0700661763548851, + 0.01789301447570324, + 0.019295183941721916, + 0.018603932112455368, + 0.0037981076166033745, + 0.01601412333548069, + 0.004603286273777485, + 0.008886762894690037, + 0.012822692282497883, + 0.006778883282095194, + 0.0042631556279957294, + 0.011038101278245449, + 0.0069572459906339645, + 0.0187078807502985, + 0.006566628813743591, + 0.017727049067616463, + 0.012835148721933365, + 0.006406029220670462, + 0.005826307460665703, + 0.0035523003898561, + 0.0032352495472878218, + 0.004285900387912989, + 0.0027285199612379074, + 0.005902206525206566, + 0.00257305265404284, + 0.0037875461857765913, + 0.0023726264480501413, + 0.002326711779460311, + 0.003849273081868887, + 0.0056209987960755825, + 0.006245593540370464, + 0.0046880715526640415, + 0.0035497411154210567, + 0.006667665671557188, + 0.005767172202467918, + 0.016551168635487556, + 0.02389640174806118, + 0.0366058349609375, + 0.0145884919911623, + 0.030158445239067078, + 0.028464097529649734, + 0.023963486775755882, + 0.01610594242811203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.01419647503644228 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073091067373753 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011073091067373753, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073091067373753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20610404014587402, + 0.07484866678714752, + 0.02525249309837818, + 0.018006769940257072, + 0.10145973414182663, + 0.09111338108778, + 0.10309673845767975, + 0.047516487538814545, + 0.09247750043869019, + 0.24012427031993866 + ], + "credit_map": [ + 0.014799561351537704, + 0.0196174718439579, + 0.0056183249689638615, + 0.015858015045523643, + 0.042639169842004776, + 0.0315561518073082, + 0.008680756203830242, + 0.06733457744121552, + 0.0067033544182777405, + 0.018351735547184944, + 0.0063518644310534, + 0.0045505911111831665, + 0.009746602736413479, + 0.011462775059044361, + 0.005850299261510372, + 0.011831444688141346, + 0.0030927821062505245, + 0.0027651458512991667, + 0.0021372116170823574, + 0.0033159528393298388, + 0.0027934564277529716, + 0.005081426352262497, + 0.001451791962608695, + 0.0046147252433001995, + 0.0017207399941980839, + 0.0035780430771410465, + 0.0023132895585149527, + 0.0014605836477130651, + 0.0031148833222687244, + 0.002074131043627858, + 0.002291092649102211, + 0.0014540051342919469, + 0.010754415765404701, + 0.006739449221640825, + 0.028790637850761414, + 0.007094763219356537, + 0.010609036311507225, + 0.026010319590568542, + 0.006434470880776644, + 0.0050266277976334095, + 0.02344881184399128, + 0.004352583084255457, + 0.016112323850393295, + 0.004133189097046852, + 0.00963610503822565, + 0.02004063129425049, + 0.00853632390499115, + 0.004853405058383942, + 0.016942622140049934, + 0.010630941949784756, + 0.014681783504784107, + 0.005253428593277931, + 0.023126395419239998, + 0.017361033707857132, + 0.007307387422770262, + 0.007793131750077009, + 0.003400128800421953, + 0.0032366577070206404, + 0.01225952710956335, + 0.005585332401096821, + 0.006034788675606251, + 0.0031506058294326067, + 0.009923819452524185, + 0.003925619181245565, + 0.005138770677149296, + 0.008230139501392841, + 0.026472702622413635, + 0.007382641546428204, + 0.008895831182599068, + 0.004825076553970575, + 0.021270841360092163, + 0.010261492803692818, + 0.022409185767173767, + 0.03181314840912819, + 0.03626025468111038, + 0.020100725814700127, + 0.036955829709768295, + 0.04170830920338631, + 0.028221627697348595, + 0.022655168548226357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.011073091067373753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, captured with the breathtaking clarity of high-frame-rate video, moves perfectly parallel to a majestic horse galloping across a vast, open plain at explosive speed. The horse, a majestic stallion with a sleek, near-black coat, epitomizes power; its long, flowing mane and wake trail behind it like silk banners in the wind. As it rumbles over the earth, every sinew and muscle group in its powerful shoulders, hips, and legs ripples visibly beneath its smooth skin. Each powerful stride is captured with surreal clarity, hooves striking the dry earth and launching in perfectly synchronized sequence, kicking up plumes of delicate golden dust that hang in the air. The entire scene is bathed in the warm, low-angle light of the afternoon sun, which creates a bright rim light around the horse's body, highlighting the texture of its coat and individual dust particles against a backdrop of sun-bleached golden grass. The atmosphere is one of uninhibited freedom and raw, breathtaking power, and the smooth, almost surreal quality of the high frame rate capture adds a graceful, balletic quality to the immense force on display.", + "seed": 30405, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_10/index_276_seed_30405_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08145414274185896, + "vlm_logprob": -0.08145414274185896, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.08145414274185896, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3054855167865753 + ], + "opposite_logprob_tokens": [ + -1.680485486984253 + ], + "logprob_reward": -0.3054855167865753, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.3054855167865753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35940021276474, + 0.05830822139978409, + 0.07985307276248932, + 0.051415883004665375, + 0.03901778906583786, + 0.04182470962405205, + 0.11193554103374481, + 0.09970059245824814, + 0.10807918757200241, + 0.05046474188566208 + ], + "credit_map": [ + 0.03424323722720146, + 0.0678841844201088, + 0.05039413645863533, + 0.03250512853264809, + 0.08957730978727341, + 0.02539132721722126, + 0.032833050936460495, + 0.02657182887196541, + 0.006480053067207336, + 0.006305623333901167, + 0.009061346761882305, + 0.008984201587736607, + 0.009311036206781864, + 0.005805693566799164, + 0.007181570399552584, + 0.005178702063858509, + 0.010987432673573494, + 0.006945360917598009, + 0.012444926425814629, + 0.007560607045888901, + 0.013624950312077999, + 0.012048282660543919, + 0.01022083219140768, + 0.006020679138600826, + 0.010265824384987354, + 0.0037669811863452196, + 0.006801316514611244, + 0.0056157601065933704, + 0.009404083713889122, + 0.004467788152396679, + 0.006260243244469166, + 0.004833889659494162, + 0.006231891922652721, + 0.003320364747196436, + 0.005426457151770592, + 0.006101297214627266, + 0.0043767960742115974, + 0.0034109277185052633, + 0.007039749063551426, + 0.003110309364274144, + 0.006937289610505104, + 0.0032891500741243362, + 0.006336485035717487, + 0.004894433077424765, + 0.005051169544458389, + 0.004414305556565523, + 0.0052102431654930115, + 0.00569163728505373, + 0.018122315406799316, + 0.00795808620750904, + 0.01749117486178875, + 0.011594228446483612, + 0.013528279960155487, + 0.014209442771971226, + 0.019101956859230995, + 0.009930059313774109, + 0.016207784414291382, + 0.00702099222689867, + 0.011627030558884144, + 0.007948954589664936, + 0.023797284811735153, + 0.014659502543509007, + 0.011429976671934128, + 0.007009072229266167, + 0.020129738375544548, + 0.008332946337759495, + 0.016784070059657097, + 0.00926508940756321, + 0.01640617474913597, + 0.009888386353850365, + 0.010124401189386845, + 0.01714838482439518, + 0.007762477733194828, + 0.004419412929564714, + 0.007346720900386572, + 0.004145618062466383, + 0.0084063820540905, + 0.0053839790634810925, + 0.008429734967648983, + 0.004570421762764454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.3054855167865753 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07011439651250839 + ], + "opposite_logprob_tokens": [ + -2.6951143741607666 + ], + "logprob_reward": -0.07011439651250839, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07011439651250839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2809051275253296, + 0.10323924571275711, + 0.04083662107586861, + 0.07685163617134094, + 0.034863024950027466, + 0.08202169835567474, + 0.0816413089632988, + 0.23576293885707855, + 0.025263043120503426, + 0.03861536458134651 + ], + "credit_map": [ + 0.03271171823143959, + 0.04858409985899925, + 0.0391678549349308, + 0.02530142292380333, + 0.062125470489263535, + 0.018286921083927155, + 0.02662377618253231, + 0.028103874996304512, + 0.012754332274198532, + 0.008095776662230492, + 0.016182083636522293, + 0.010595610365271568, + 0.013021436519920826, + 0.009096524678170681, + 0.020550359040498734, + 0.012943126261234283, + 0.006904255598783493, + 0.003525074804201722, + 0.006931046023964882, + 0.004166190046817064, + 0.006383625324815512, + 0.0048613399267196655, + 0.004991957917809486, + 0.003073134459555149, + 0.017790446057915688, + 0.0051399702206254005, + 0.011011575348675251, + 0.008056175895035267, + 0.013565931469202042, + 0.007710670121014118, + 0.008038556203246117, + 0.005538302473723888, + 0.004950476810336113, + 0.0027389272581785917, + 0.005023641046136618, + 0.0046443454921245575, + 0.004940750543028116, + 0.003469756804406643, + 0.0057665701024234295, + 0.0033285559620708227, + 0.01848425716161728, + 0.005858822725713253, + 0.013376225717365742, + 0.006480303127318621, + 0.015440206043422222, + 0.007174776401370764, + 0.009777977131307125, + 0.0054291314445436, + 0.018897967413067818, + 0.00537438178434968, + 0.013910514302551746, + 0.005837102886289358, + 0.010910606011748314, + 0.007899992167949677, + 0.012498130090534687, + 0.006312609650194645, + 0.03928341343998909, + 0.016901614144444466, + 0.0341469869017601, + 0.019053157418966293, + 0.04695805534720421, + 0.029624607414007187, + 0.026569772511720657, + 0.0232253260910511, + 0.002981761237606406, + 0.00328843854367733, + 0.003207579953595996, + 0.003167536109685898, + 0.00386737659573555, + 0.002155442489311099, + 0.0028813292738050222, + 0.0037135782185941935, + 0.008809625171124935, + 0.0031158095225691795, + 0.006609322968870401, + 0.002826192881911993, + 0.004040352534502745, + 0.0029253270477056503, + 0.008425927720963955, + 0.001862805336713791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.07011439651250839 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747319363057613 + ], + "opposite_logprob_tokens": [ + -5.006747245788574 + ], + "logprob_reward": -0.006747319363057613, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747319363057613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3462217152118683, + 0.10572139918804169, + 0.05687616392970085, + 0.051167938858270645, + 0.035159360617399216, + 0.0513361319899559, + 0.13359950482845306, + 0.1709083616733551, + 0.030306188389658928, + 0.018703224137425423 + ], + "credit_map": [ + 0.04058585688471794, + 0.058087851852178574, + 0.046612050384283066, + 0.032067395746707916, + 0.07459291815757751, + 0.02406427450478077, + 0.03302755579352379, + 0.03718382865190506, + 0.01394547801464796, + 0.00852182786911726, + 0.017831461504101753, + 0.01002022810280323, + 0.013782767578959465, + 0.007888698019087315, + 0.022513410076498985, + 0.011217533610761166, + 0.01324682030826807, + 0.004294419661164284, + 0.009145934134721756, + 0.006699895020574331, + 0.006494942121207714, + 0.00646980619058013, + 0.006680690683424473, + 0.0038436581380665302, + 0.012316315434873104, + 0.0035836303140968084, + 0.007137757260352373, + 0.00528932549059391, + 0.008019494824111462, + 0.005614605266600847, + 0.005618350580334663, + 0.0035884648095816374, + 0.004770240746438503, + 0.0023239830043166876, + 0.004407238680869341, + 0.004045349080115557, + 0.005795513745397329, + 0.0038203373551368713, + 0.006305201444774866, + 0.0036914933007210493, + 0.007756272330880165, + 0.00509537011384964, + 0.0073101455345749855, + 0.004946552217006683, + 0.007657269015908241, + 0.005748845171183348, + 0.006808541715145111, + 0.006013140082359314, + 0.017230484634637833, + 0.01049327477812767, + 0.021746689453721046, + 0.009111286140978336, + 0.018192410469055176, + 0.023570634424686432, + 0.014097784645855427, + 0.01915694773197174, + 0.027149459347128868, + 0.01291641779243946, + 0.024061281234025955, + 0.01367352344095707, + 0.03622293099761009, + 0.021831272169947624, + 0.01992000639438629, + 0.015133469365537167, + 0.004665244370698929, + 0.0029295028652995825, + 0.0048560043796896935, + 0.002284659305587411, + 0.004762313328683376, + 0.0027292445302009583, + 0.0029775002039968967, + 0.0051017203368246555, + 0.002629529917612672, + 0.002122976118698716, + 0.003419561078771949, + 0.0019236154621466994, + 0.002511603059247136, + 0.002298639854416251, + 0.0017574006924405694, + 0.002039897721260786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.006747319363057613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818835735321045 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818835735321045, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818835735321045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26544076204299927, + 0.1523047536611557, + 0.12149189412593842, + 0.13207657635211945, + 0.07792042940855026, + 0.07253208011388779, + 0.038514796644449234, + 0.06602640450000763, + 0.03481049835681915, + 0.038881830871105194 + ], + "credit_map": [ + 0.028295883908867836, + 0.044888343662023544, + 0.038826730102300644, + 0.02848931960761547, + 0.05970613658428192, + 0.0186789408326149, + 0.02688457816839218, + 0.019670795649290085, + 0.016129516065120697, + 0.010902057401835918, + 0.025074267759919167, + 0.01607009582221508, + 0.02252003364264965, + 0.012822088785469532, + 0.02831488847732544, + 0.02047179639339447, + 0.02052905224263668, + 0.009649066254496574, + 0.016899364069104195, + 0.013290971517562866, + 0.018555806949734688, + 0.01601416990160942, + 0.016321761533617973, + 0.010231698863208294, + 0.03505677729845047, + 0.009096202440559864, + 0.020432326942682266, + 0.015126774087548256, + 0.018106894567608833, + 0.012456780299544334, + 0.012725009582936764, + 0.009075801819562912, + 0.015816064551472664, + 0.0052713206969201565, + 0.011763560585677624, + 0.007364874705672264, + 0.010150806047022343, + 0.007811518851667643, + 0.012229151092469692, + 0.0075131370685994625, + 0.014805087819695473, + 0.004632615018635988, + 0.01105575542896986, + 0.007076702546328306, + 0.0130948256701231, + 0.006687662098556757, + 0.00935207586735487, + 0.00582735612988472, + 0.005954680033028126, + 0.0027914969250559807, + 0.005902090109884739, + 0.0025690963957458735, + 0.004589679650962353, + 0.005940109491348267, + 0.006332080811262131, + 0.004435566253960133, + 0.010794776491820812, + 0.005186164285987616, + 0.008940468542277813, + 0.005953246261924505, + 0.012241385877132416, + 0.00897526927292347, + 0.009534814395010471, + 0.004400278441607952, + 0.004395379684865475, + 0.004506736993789673, + 0.004420559387654066, + 0.003275482915341854, + 0.0062323156744241714, + 0.003419126383960247, + 0.004167104605585337, + 0.004393795505166054, + 0.007123490795493126, + 0.002923237392678857, + 0.006344377528876066, + 0.0035695850383490324, + 0.004502494353801012, + 0.0031307071913033724, + 0.008737051859498024, + 0.0025508867111057043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01818835735321045 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle relative to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera moving parallel to a horse galloping across an open plain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horse depicted as a stallion with a near-black coat and a long, flowing mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible plumes of golden dust kicked up by the horse's hooves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm afternoon sunlight creating a bright rim light around the horse?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735123693943024 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735123693943024, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735123693943024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18773141503334045, + 0.040220681577920914, + 0.1430320292711258, + 0.07226132601499557, + 0.03684563562273979, + 0.06653863191604614, + 0.07123655080795288, + 0.20592066645622253, + 0.12789782881736755, + 0.04831529036164284 + ], + "credit_map": [ + 0.02197943441569805, + 0.02684907242655754, + 0.02885003201663494, + 0.017914023250341415, + 0.038202621042728424, + 0.012950385920703411, + 0.019527370110154152, + 0.02145845629274845, + 0.005396685563027859, + 0.0033400743268430233, + 0.006250154227018356, + 0.004914510529488325, + 0.006527937017381191, + 0.004649673588573933, + 0.0047663552686572075, + 0.0043752845376729965, + 0.03345668688416481, + 0.011120473966002464, + 0.02506457082927227, + 0.016340168192982674, + 0.019032850861549377, + 0.010602095164358616, + 0.017684563994407654, + 0.009730597026646137, + 0.01360414270311594, + 0.004817485343664885, + 0.010471305809915066, + 0.007638019509613514, + 0.014185535721480846, + 0.007919564843177795, + 0.008012651465833187, + 0.005612608976662159, + 0.006860140711069107, + 0.0024539714213460684, + 0.005220017395913601, + 0.0032566492445766926, + 0.003086461452767253, + 0.003888951148837805, + 0.007077524904161692, + 0.005001916084438562, + 0.011486141942441463, + 0.004379228688776493, + 0.010989475063979626, + 0.005968510173261166, + 0.010955819860100746, + 0.0068677328526973724, + 0.008819608017802238, + 0.007072107400745153, + 0.009615974500775337, + 0.004687597043812275, + 0.010776231065392494, + 0.0074334521777927876, + 0.009530509822070599, + 0.008151035755872726, + 0.014785382896661758, + 0.006256360560655594, + 0.04069555178284645, + 0.015851598232984543, + 0.0371260866522789, + 0.01674824021756649, + 0.030093839392066002, + 0.023860488086938858, + 0.02457134611904621, + 0.01697349175810814, + 0.0211932510137558, + 0.013463964685797691, + 0.020154383033514023, + 0.012014031410217285, + 0.02048059180378914, + 0.010283923707902431, + 0.015223299153149128, + 0.015084376558661461, + 0.009564300999045372, + 0.003445821348577738, + 0.00800622720271349, + 0.004230121150612831, + 0.005509857088327408, + 0.0034975814633071423, + 0.010723721235990524, + 0.0033376545179635286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.006735123693943024 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_11/index_tensor([5889, 5889])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_11/index_tensor([5889, 5889])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6353c552db4e150041c9fb3d1578ae994d32625c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_11/index_tensor([5889, 5889])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot captured from a level angle, focuses squarely on the face of a middle-aged man, his brow furrowed in deep thought over a vexing problem. A single, warm light source, like the light in a Rembrandt painting, dramatically sculpts his features, illuminating one cheek and making tiny beads of sweat glisten on his temple, while the other side is shrouded in deep shadow. Deep vertical wrinkles form between his brows from intense concentration, while horizontal lines of thought crisscross his forehead. His eyes are slightly narrowed, his gaze unfocused, as if gazing into the distance through the lens. His lips are pressed into a firm, resolute line, his jaw muscles subtly tensed, betraying an inner struggle. A few strands of dark, slightly damp hair fall across his forehead, the texture and pores of his skin clearly visible in the lens, while the blurred background suggests a late-night study. The entire picture is shrouded in an intense silence, creating an atmosphere of extreme concentration, tension and inner struggle, allowing the audience to directly feel the intense thinking going on in his mind.", + "seed": 26221, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_11/index_5889_seed_26221_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05234641821589321, + "vlm_logprob": -0.05234641821589321, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.05234641821589321, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022885605692863464 + ], + "opposite_logprob_tokens": [ + -4.022885799407959 + ], + "logprob_reward": -0.022885605692863464, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.022885605692863464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10942313820123672, + 0.13859868049621582, + 0.27552419900894165, + 0.06835923343896866, + 0.03268913924694061, + 0.04643953964114189, + 0.09122597426176071, + 0.09602627903223038, + 0.058330513536930084, + 0.08338330686092377 + ], + "credit_map": [ + 0.008073750883340836, + 0.009284359402954578, + 0.009399590082466602, + 0.013172068633139133, + 0.020926574245095253, + 0.01979336328804493, + 0.017308641225099564, + 0.011464791372418404, + 0.011152755469083786, + 0.01662522181868553, + 0.012890954501926899, + 0.01816706918179989, + 0.022114943712949753, + 0.019707653671503067, + 0.016491562128067017, + 0.02144850790500641, + 0.021808134391903877, + 0.03708607703447342, + 0.023493383079767227, + 0.032855939120054245, + 0.04700978845357895, + 0.03843018040060997, + 0.0472554974257946, + 0.027585186064243317, + 0.005514603573828936, + 0.008842728100717068, + 0.004931577946990728, + 0.008198981173336506, + 0.012214028276503086, + 0.012029275298118591, + 0.010016598738729954, + 0.0066114384680986404, + 0.002889966359362006, + 0.0034707870800048113, + 0.0027703153900802135, + 0.003612630534917116, + 0.005202129483222961, + 0.00597588624805212, + 0.004622182343155146, + 0.004145238548517227, + 0.004032526630908251, + 0.0048796809278428555, + 0.004598911385983229, + 0.004912436008453369, + 0.006091933697462082, + 0.009183340705931187, + 0.007544228807091713, + 0.005196473561227322, + 0.008315421640872955, + 0.007028491236269474, + 0.006548333913087845, + 0.009085973724722862, + 0.017454545944929123, + 0.01931016705930233, + 0.011474466882646084, + 0.012008567340672016, + 0.005724817980080843, + 0.0068705216981470585, + 0.004806870128959417, + 0.010933641344308853, + 0.02380136400461197, + 0.019664732739329338, + 0.016248438507318497, + 0.00797587726265192, + 0.0051070209592580795, + 0.007383142132312059, + 0.004902295768260956, + 0.0063522737473249435, + 0.012063086032867432, + 0.008972849696874619, + 0.00860391091555357, + 0.004945926833897829, + 0.0075078909285366535, + 0.006828168407082558, + 0.007490815594792366, + 0.0072826361283659935, + 0.009607275947928429, + 0.023580215871334076, + 0.009651661850512028, + 0.011434642598032951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.022885605692863464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869326949119568 + ], + "opposite_logprob_tokens": [ + -3.0486931800842285 + ], + "logprob_reward": -0.04869326949119568, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869326949119568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23481327295303345, + 0.059174153953790665, + 0.0582350455224514, + 0.0767662301659584, + 0.06972382962703705, + 0.09106536954641342, + 0.1560724526643753, + 0.13071413338184357, + 0.02948775701224804, + 0.09394776821136475 + ], + "credit_map": [ + 0.01894901692867279, + 0.02009766362607479, + 0.019868439063429832, + 0.025912215933203697, + 0.028767138719558716, + 0.05600018799304962, + 0.042270414531230927, + 0.02294819988310337, + 0.0035535250790417194, + 0.004217596724629402, + 0.0045104920864105225, + 0.004599284380674362, + 0.008629649877548218, + 0.010974946431815624, + 0.011282937601208687, + 0.011405723169445992, + 0.0042073591612279415, + 0.004204472992569208, + 0.004735772032290697, + 0.004271803889423609, + 0.009703489951789379, + 0.010022313334047794, + 0.010539638809859753, + 0.010550194419920444, + 0.005344097036868334, + 0.009309263899922371, + 0.004394453018903732, + 0.008407370187342167, + 0.016789348796010017, + 0.010629992932081223, + 0.012204890139400959, + 0.00968681275844574, + 0.004842563532292843, + 0.00879568699747324, + 0.004337816499173641, + 0.008311078883707523, + 0.013410345651209354, + 0.013439356349408627, + 0.009711840189993382, + 0.0068751429207623005, + 0.010263141244649887, + 0.0108665581792593, + 0.008171546272933483, + 0.010897836647927761, + 0.01193949580192566, + 0.01705431379377842, + 0.0117822689935565, + 0.010090204887092113, + 0.013807566836476326, + 0.011781660839915276, + 0.014792525209486485, + 0.016468998044729233, + 0.02151467651128769, + 0.036773305386304855, + 0.020467156544327736, + 0.02046656794846058, + 0.005287016276270151, + 0.008279831148684025, + 0.005764619447290897, + 0.007300328463315964, + 0.04045780003070831, + 0.030429743230342865, + 0.01817346177995205, + 0.015021327883005142, + 0.0023884610272943974, + 0.002060827799141407, + 0.0031473326962441206, + 0.0022828697692602873, + 0.006061990279704332, + 0.005155822727829218, + 0.005917363800108433, + 0.0024730891454964876, + 0.007493340410292149, + 0.008210103958845139, + 0.009023867547512054, + 0.011204134672880173, + 0.018518619239330292, + 0.014344621449708939, + 0.01333122979849577, + 0.01182185485959053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.04869326949119568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015237637562677264 + ], + "opposite_logprob_tokens": [ + -6.501523971557617 + ], + "logprob_reward": -0.0015237637562677264, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015237637562677264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22669066488742828, + 0.05780410394072533, + 0.11935479193925858, + 0.048925504088401794, + 0.053922995924949646, + 0.0745166540145874, + 0.09314961731433868, + 0.22099217772483826, + 0.0576472245156765, + 0.04699629917740822 + ], + "credit_map": [ + 0.017811071127653122, + 0.024222301319241524, + 0.016330264508724213, + 0.025171823799610138, + 0.050041019916534424, + 0.03788355365395546, + 0.02723788283765316, + 0.027992727234959602, + 0.0037131605204194784, + 0.0046360185369849205, + 0.0039943791925907135, + 0.005068272352218628, + 0.00791554618626833, + 0.009710405021905899, + 0.010721754282712936, + 0.012044559232890606, + 0.008476133458316326, + 0.011050854809582233, + 0.00834977813065052, + 0.011414535343647003, + 0.015432064421474934, + 0.01759421080350876, + 0.023285094648599625, + 0.023752115666866302, + 0.00292400224134326, + 0.004325576592236757, + 0.0026095681823790073, + 0.003900512121617794, + 0.011411710642278194, + 0.0071572307497262955, + 0.00947536900639534, + 0.007121531292796135, + 0.004377278499305248, + 0.006291709374636412, + 0.004241232294589281, + 0.005138331092894077, + 0.007179714739322662, + 0.009897408075630665, + 0.010460181161761284, + 0.0063371360301971436, + 0.006049168296158314, + 0.008672679774463177, + 0.005190310068428516, + 0.006475526839494705, + 0.010525164194405079, + 0.0160239115357399, + 0.013631152920424938, + 0.007948744110763073, + 0.0082430774345994, + 0.008758963085711002, + 0.009361425414681435, + 0.012420511804521084, + 0.012848756276071072, + 0.018168935552239418, + 0.011640074662864208, + 0.011707860976457596, + 0.010441425256431103, + 0.0160811860114336, + 0.00980351958423853, + 0.019823376089334488, + 0.06798949092626572, + 0.04373207315802574, + 0.034653373062610626, + 0.018467729911208153, + 0.004610343370586634, + 0.0035480009391903877, + 0.003876510774716735, + 0.004421984311193228, + 0.010233910754323006, + 0.015882141888141632, + 0.008220506832003593, + 0.00685381842777133, + 0.0033969355281442404, + 0.0043114484287798405, + 0.00432501919567585, + 0.006451056804507971, + 0.007501847110688686, + 0.008653019554913044, + 0.006399914622306824, + 0.005957054439932108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0015237637562677264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653167635202408 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.008653167635202408, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653167635202408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23719139397144318, + 0.11090350896120071, + 0.12488409876823425, + 0.05194491147994995, + 0.021402234211564064, + 0.07519586384296417, + 0.1458597332239151, + 0.07468462735414505, + 0.06964819133281708, + 0.08828537911176682 + ], + "credit_map": [ + 0.019122052937746048, + 0.019283069297671318, + 0.014447632245719433, + 0.022208113223314285, + 0.059064432978630066, + 0.03677462413907051, + 0.04001762345433235, + 0.026273848488926888, + 0.010053587146103382, + 0.011883670464158058, + 0.011070276610553265, + 0.009964882396161556, + 0.01625092327594757, + 0.023536914959549904, + 0.016569077968597412, + 0.011574171483516693, + 0.009557037614285946, + 0.0166004691272974, + 0.008381097577512264, + 0.013624770566821098, + 0.01824096404016018, + 0.022164098918437958, + 0.02042558789253235, + 0.01589006930589676, + 0.003991588018834591, + 0.006242561619728804, + 0.003675145097076893, + 0.007540035527199507, + 0.00846710056066513, + 0.009071916341781616, + 0.0077479989267885685, + 0.0052085635252296925, + 0.00156915164552629, + 0.002867767121642828, + 0.0018788964953273535, + 0.002348256064578891, + 0.0030544803012162447, + 0.003420509397983551, + 0.0037014258559793234, + 0.0025617466308176517, + 0.0067644277587533, + 0.00876876711845398, + 0.005972260609269142, + 0.008019926957786083, + 0.015442181378602982, + 0.012486160732805729, + 0.010696418583393097, + 0.00704572256654501, + 0.009397896006703377, + 0.012447075918316841, + 0.012226554565131664, + 0.017656557261943817, + 0.03531511500477791, + 0.02603594772517681, + 0.019502321258187294, + 0.01327827200293541, + 0.006313747260719538, + 0.004996378906071186, + 0.005190268624573946, + 0.008854633197188377, + 0.009695551358163357, + 0.019157301634550095, + 0.010379476472735405, + 0.010097265243530273, + 0.005948004312813282, + 0.004303023684769869, + 0.006372981239110231, + 0.005856257863342762, + 0.01075578760355711, + 0.017888855189085007, + 0.011121267452836037, + 0.007402009330689907, + 0.006485256832093, + 0.005333778914064169, + 0.008686116896569729, + 0.010148376226425171, + 0.011443772353231907, + 0.023589979857206345, + 0.01140813808888197, + 0.011189962737262249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.008653167635202408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17997628450393677 + ], + "opposite_logprob_tokens": [ + -1.804976224899292 + ], + "logprob_reward": -0.17997628450393677, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17997628450393677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30718186497688293, + 0.14314627647399902, + 0.05348706245422363, + 0.07124149054288864, + 0.08845708519220352, + 0.0432671457529068, + 0.11076829582452774, + 0.08182979375123978, + 0.047661907970905304, + 0.05295904725790024 + ], + "credit_map": [ + 0.017522057518363, + 0.02427641488611698, + 0.019404323771595955, + 0.021952850744128227, + 0.07447713613510132, + 0.04608694463968277, + 0.06212323158979416, + 0.041338901966810226, + 0.010282959789037704, + 0.019160930067300797, + 0.010536578483879566, + 0.01745939627289772, + 0.02663968876004219, + 0.01926729828119278, + 0.022988436743617058, + 0.016810990869998932, + 0.004009673371911049, + 0.0046949125826358795, + 0.0049051037058234215, + 0.006308657582849264, + 0.007081531919538975, + 0.009983747266232967, + 0.0076645235531032085, + 0.008838915266096592, + 0.004313250072300434, + 0.006329105235636234, + 0.0050274101085960865, + 0.005094071850180626, + 0.01638174243271351, + 0.010552494786679745, + 0.012869095429778099, + 0.010674322955310345, + 0.006247981917113066, + 0.007800435181707144, + 0.006482376251369715, + 0.010526025667786598, + 0.012883240357041359, + 0.013770746067166328, + 0.013563326559960842, + 0.0171829666942358, + 0.00368527602404356, + 0.004813459236174822, + 0.0038981479592621326, + 0.003546979743987322, + 0.00690377876162529, + 0.008470254950225353, + 0.0061310408636927605, + 0.005818210542201996, + 0.00676730927079916, + 0.009462857618927956, + 0.007820322178304195, + 0.012497758492827415, + 0.028662919998168945, + 0.020775018259882927, + 0.015645181760191917, + 0.009136933833360672, + 0.003408820368349552, + 0.004686100408434868, + 0.004296040162444115, + 0.005941852927207947, + 0.023184165358543396, + 0.01428245659917593, + 0.016165174543857574, + 0.009865188039839268, + 0.0034640023950487375, + 0.002740316791459918, + 0.002991665853187442, + 0.003821304300799966, + 0.008867509663105011, + 0.012656708247959614, + 0.00820664782077074, + 0.004913756158202887, + 0.004024846479296684, + 0.004855226259678602, + 0.0070295860059559345, + 0.008954888209700584, + 0.008485781960189342, + 0.008450021035969257, + 0.006854098290205002, + 0.004304599482566118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.17997628450393677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot captured from a level angle, focuses squarely on the face of a middle-aged man, his brow furrowed in deep thought over a vexing problem. A single, warm light source, like the light in a Rembrandt painting, dramatically sculpts his features, illuminating one cheek and making tiny beads of sweat glisten on his temple, while the other side is shrouded in deep shadow. Deep vertical wrinkles form between his brows from intense concentration, while horizontal lines of thought crisscross his forehead. His eyes are slightly narrowed, his gaze unfocused, as if gazing into the distance through the lens. His lips are pressed into a firm, resolute line, his jaw muscles subtly tensed, betraying an inner struggle. A few strands of dark, slightly damp hair fall across his forehead, the texture and pores of his skin clearly visible in the lens, while the blurred background suggests a late-night study. The entire picture is shrouded in an intense silence, creating an atmosphere of extreme concentration, tension and inner struggle, allowing the audience to directly feel the intense thinking going on in his mind.", + "seed": 26222, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_11/index_5889_seed_26222_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.029043900500983, + "vlm_logprob": -0.029043900500983, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.029043900500983, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005140540190041065 + ], + "opposite_logprob_tokens": [ + -6.50514030456543 + ], + "logprob_reward": -0.005140540190041065, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.005140540190041065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2099030464887619, + 0.07310052216053009, + 0.14163687825202942, + 0.0652267187833786, + 0.08229860663414001, + 0.0663062259554863, + 0.08708883076906204, + 0.0730244517326355, + 0.09198527783155441, + 0.10942943394184113 + ], + "credit_map": [ + 0.021136291325092316, + 0.02760181948542595, + 0.02030785195529461, + 0.02338993363082409, + 0.025346551090478897, + 0.03105134516954422, + 0.03472593426704407, + 0.02634333074092865, + 0.006269108969718218, + 0.0098753422498703, + 0.007523097097873688, + 0.008158043026924133, + 0.010314065963029861, + 0.010186486877501011, + 0.01343559380620718, + 0.007338791154325008, + 0.011372940614819527, + 0.03218812495470047, + 0.011122487485408783, + 0.020383168011903763, + 0.015519501641392708, + 0.01693757250905037, + 0.01698021963238716, + 0.017132865265011787, + 0.006917500868439674, + 0.008480396121740341, + 0.00656764954328537, + 0.005809135735034943, + 0.008403115905821323, + 0.008508270606398582, + 0.011290892958641052, + 0.009249753318727016, + 0.0077825565822422504, + 0.010697956196963787, + 0.008140903897583485, + 0.009372909553349018, + 0.011215361766517162, + 0.011857764795422554, + 0.013962960802018642, + 0.009268191643059254, + 0.005881162825971842, + 0.008800269104540348, + 0.006279855500906706, + 0.009368406608700752, + 0.00880702119320631, + 0.009151463396847248, + 0.01273560244590044, + 0.005282446276396513, + 0.010967937298119068, + 0.009438326582312584, + 0.00936481636017561, + 0.00904131680727005, + 0.010162795893847942, + 0.01215486228466034, + 0.011937787756323814, + 0.01402098685503006, + 0.006691409274935722, + 0.012708221562206745, + 0.006655780132859945, + 0.008252164348959923, + 0.007482671644538641, + 0.01058944221585989, + 0.013039578683674335, + 0.007605188526213169, + 0.007050340063869953, + 0.011193443089723587, + 0.008116357028484344, + 0.009737182408571243, + 0.016025831922888756, + 0.0143650583922863, + 0.016115667298436165, + 0.009381400421261787, + 0.013639334589242935, + 0.019591396674513817, + 0.012459388934075832, + 0.014225944876670837, + 0.01382511667907238, + 0.014240597374737263, + 0.012270959094166756, + 0.009176686406135559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.005140540190041065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373133063316345 + ], + "opposite_logprob_tokens": [ + -3.6513731479644775 + ], + "logprob_reward": -0.026373133063316345, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373133063316345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1489151120185852, + 0.12341169267892838, + 0.065837562084198, + 0.08503273874521255, + 0.1416488140821457, + 0.09046731889247894, + 0.1542523354291916, + 0.06581874936819077, + 0.04813048243522644, + 0.07648520171642303 + ], + "credit_map": [ + 0.021486273035407066, + 0.018806014209985733, + 0.016108183190226555, + 0.016276374459266663, + 0.01781659573316574, + 0.024528399109840393, + 0.017535192891955376, + 0.016358062624931335, + 0.010399574413895607, + 0.019251959398388863, + 0.009009636007249355, + 0.012228650040924549, + 0.01878644898533821, + 0.016019629314541817, + 0.01913706213235855, + 0.018578730523586273, + 0.006350071635097265, + 0.015871040523052216, + 0.004568557720631361, + 0.008793160319328308, + 0.008419212885200977, + 0.007335114758461714, + 0.00775989331305027, + 0.006740509532392025, + 0.011246845126152039, + 0.0056269289925694466, + 0.010446503758430481, + 0.008688715286552906, + 0.008498220704495907, + 0.016480702906847, + 0.010528074577450752, + 0.013516746461391449, + 0.012972716242074966, + 0.021853696554899216, + 0.01249268464744091, + 0.018578024581074715, + 0.014039941132068634, + 0.01742708496749401, + 0.022595586255192757, + 0.021689075976610184, + 0.010916336439549923, + 0.010352560319006443, + 0.010099321603775024, + 0.012514431960880756, + 0.01114202942699194, + 0.011389209888875484, + 0.013443262316286564, + 0.010610152967274189, + 0.016621259972453117, + 0.030720245093107224, + 0.014095085673034191, + 0.017605511471629143, + 0.016575483605265617, + 0.018608922138810158, + 0.01834067702293396, + 0.021685132756829262, + 0.008461354300379753, + 0.013724839314818382, + 0.005814999341964722, + 0.007022978272289038, + 0.004987907130271196, + 0.008145160973072052, + 0.008786708116531372, + 0.008874798193573952, + 0.005760666448622942, + 0.004710615612566471, + 0.007110013160854578, + 0.0064286235719919205, + 0.005032781977206469, + 0.0069485437124967575, + 0.006621977314352989, + 0.005517258774489164, + 0.008023458532989025, + 0.018057342618703842, + 0.007565556559711695, + 0.0069722505286335945, + 0.010469579137861729, + 0.009139209985733032, + 0.009760658256709576, + 0.006497138645499945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.026373133063316345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645151551812887 + ], + "opposite_logprob_tokens": [ + -5.379645347595215 + ], + "logprob_reward": -0.004645151551812887, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645151551812887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10290940850973129, + 0.0588236004114151, + 0.07164496183395386, + 0.16181492805480957, + 0.06079617887735367, + 0.19243475794792175, + 0.08521956205368042, + 0.03968583419919014, + 0.06997343897819519, + 0.1566973328590393 + ], + "credit_map": [ + 0.014743020758032799, + 0.015254680998623371, + 0.01028390321880579, + 0.011261701583862305, + 0.008161494508385658, + 0.013121822848916054, + 0.01589217409491539, + 0.0141906151548028, + 0.0061041442677378654, + 0.011469629593193531, + 0.006360921543091536, + 0.005887965206056833, + 0.00802743248641491, + 0.007551874499768019, + 0.0076562101021409035, + 0.005765421781688929, + 0.006547349039465189, + 0.005984494928270578, + 0.007278345990926027, + 0.011086620390415192, + 0.007627489045262337, + 0.00968958344310522, + 0.014917898923158646, + 0.008513182401657104, + 0.02264832891523838, + 0.014135715551674366, + 0.02226804941892624, + 0.019052715972065926, + 0.013639123179018497, + 0.025245781987905502, + 0.017070099711418152, + 0.027755100280046463, + 0.007057773880660534, + 0.005920084659010172, + 0.0071476479060947895, + 0.005069181323051453, + 0.00778944930061698, + 0.0088820094242692, + 0.010357155464589596, + 0.008572877384722233, + 0.025893084704875946, + 0.021164417266845703, + 0.01897629350423813, + 0.022986313328146935, + 0.018645504489541054, + 0.027592092752456665, + 0.02511785551905632, + 0.0320592038333416, + 0.009246601723134518, + 0.014440679922699928, + 0.01109741348773241, + 0.008645952679216862, + 0.008987401612102985, + 0.011484509333968163, + 0.012308637611567974, + 0.009008369408547878, + 0.00490944879129529, + 0.004016811493784189, + 0.00454874848946929, + 0.003855533665046096, + 0.0035388884134590626, + 0.00646126177161932, + 0.0059251622296869755, + 0.006429979577660561, + 0.008313244208693504, + 0.011440836824476719, + 0.00782678835093975, + 0.007253408897668123, + 0.00692156795412302, + 0.009546828456223011, + 0.006077832076698542, + 0.012592939659953117, + 0.014760888181626797, + 0.022324077785015106, + 0.029110245406627655, + 0.015209231525659561, + 0.022206982597708702, + 0.019805533811450005, + 0.0192854106426239, + 0.013994960114359856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.004645151551812887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649858646094799 + ], + "opposite_logprob_tokens": [ + -4.758649826049805 + ], + "logprob_reward": -0.008649858646094799, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649858646094799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2517119348049164, + 0.12688711285591125, + 0.11299837380647659, + 0.10118239372968674, + 0.07242056727409363, + 0.10311068594455719, + 0.07138556241989136, + 0.053475916385650635, + 0.03172669559717178, + 0.07510071247816086 + ], + "credit_map": [ + 0.028236377984285355, + 0.035585176199674606, + 0.02576833963394165, + 0.028744228184223175, + 0.030206535011529922, + 0.03987125679850578, + 0.03746914491057396, + 0.02583090402185917, + 0.012380369938910007, + 0.017870813608169556, + 0.01159767434000969, + 0.014097839593887329, + 0.015260379761457443, + 0.01918482407927513, + 0.01990179717540741, + 0.01659342087805271, + 0.0076988451182842255, + 0.020958928391337395, + 0.010207685641944408, + 0.015670785680413246, + 0.014873472042381763, + 0.014404673129320145, + 0.018055519089102745, + 0.011128471232950687, + 0.006616758648306131, + 0.012565338052809238, + 0.009822174906730652, + 0.013912211172282696, + 0.010321413166821003, + 0.0128836864605546, + 0.024092353880405426, + 0.010968470014631748, + 0.007049018982797861, + 0.01312555093318224, + 0.006869879551231861, + 0.00906202383339405, + 0.008270449936389923, + 0.010193015448749065, + 0.01125322375446558, + 0.006597406696528196, + 0.010768800042569637, + 0.011380365118384361, + 0.010466307401657104, + 0.011644132435321808, + 0.014344324357807636, + 0.014840709045529366, + 0.01762288063764572, + 0.012043170630931854, + 0.009657551534473896, + 0.004671594128012657, + 0.012757343240082264, + 0.0062277489341795444, + 0.006312579847872257, + 0.011278611607849598, + 0.010489269159734249, + 0.00999087281525135, + 0.00387139106169343, + 0.004162058234214783, + 0.005650484934449196, + 0.005536105949431658, + 0.007028499618172646, + 0.008126932196319103, + 0.012136565521359444, + 0.0069638811983168125, + 0.004286786541342735, + 0.0028867973014712334, + 0.0032704805489629507, + 0.0040008812211453915, + 0.0030308279674500227, + 0.004464789293706417, + 0.005550800357013941, + 0.004235335625708103, + 0.007496439851820469, + 0.012748387642204762, + 0.009628611616790295, + 0.008828884921967983, + 0.009143052622675896, + 0.00845393817871809, + 0.008556731045246124, + 0.010244673117995262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008649858646094799 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1004108190536499 + ], + "opposite_logprob_tokens": [ + -2.3504109382629395 + ], + "logprob_reward": -0.1004108190536499, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1004108190536499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06583239138126373, + 0.08723097294569016, + 0.047288745641708374, + 0.10327022522687912, + 0.09844052046537399, + 0.0453227199614048, + 0.05323856323957443, + 0.09347036480903625, + 0.10950411856174469, + 0.29640141129493713 + ], + "credit_map": [ + 0.00901858601719141, + 0.01112334430217743, + 0.007189874537289143, + 0.007466586772352457, + 0.0071693905629217625, + 0.011665127240121365, + 0.007271372247487307, + 0.004928112030029297, + 0.010431974194943905, + 0.014108533039689064, + 0.009416742250323296, + 0.00851339753717184, + 0.010179431177675724, + 0.01126241497695446, + 0.013946468941867352, + 0.009372008964419365, + 0.005432127509266138, + 0.00786327663809061, + 0.006122733931988478, + 0.00485956110060215, + 0.005248646251857281, + 0.007716737687587738, + 0.003798613091930747, + 0.006247049663215876, + 0.010998913086950779, + 0.009103426709771156, + 0.015248243696987629, + 0.013207769021391869, + 0.008674075827002525, + 0.013487730175256729, + 0.018920572474598885, + 0.01362949050962925, + 0.014609424397349358, + 0.012341215275228024, + 0.012617949396371841, + 0.009312817826867104, + 0.009436558932065964, + 0.012261440977454185, + 0.014220925979316235, + 0.01364019513130188, + 0.006229588761925697, + 0.004742185119539499, + 0.004955227021127939, + 0.005219226237386465, + 0.004837981425225735, + 0.00657859118655324, + 0.00541809294372797, + 0.007341830525547266, + 0.0047559794038534164, + 0.0047601088881492615, + 0.0051332032307982445, + 0.006809378042817116, + 0.005167876835912466, + 0.0086669297888875, + 0.01026037335395813, + 0.007684711366891861, + 0.010731091722846031, + 0.0159749872982502, + 0.00975762028247118, + 0.010514357127249241, + 0.010903121903538704, + 0.013408144004642963, + 0.013482211157679558, + 0.008698827587068081, + 0.01107195857912302, + 0.015821482986211777, + 0.014080419205129147, + 0.01235636230558157, + 0.012053893879055977, + 0.014382023364305496, + 0.011612875387072563, + 0.018125101923942566, + 0.03044445440173149, + 0.04152991622686386, + 0.05511211231350899, + 0.028609557077288628, + 0.03813322260975838, + 0.04043643921613693, + 0.03707699105143547, + 0.025058718398213387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.1004108190536499 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot captured from a level angle, focuses squarely on the face of a middle-aged man, his brow furrowed in deep thought over a vexing problem. A single, warm light source, like the light in a Rembrandt painting, dramatically sculpts his features, illuminating one cheek and making tiny beads of sweat glisten on his temple, while the other side is shrouded in deep shadow. Deep vertical wrinkles form between his brows from intense concentration, while horizontal lines of thought crisscross his forehead. His eyes are slightly narrowed, his gaze unfocused, as if gazing into the distance through the lens. His lips are pressed into a firm, resolute line, his jaw muscles subtly tensed, betraying an inner struggle. A few strands of dark, slightly damp hair fall across his forehead, the texture and pores of his skin clearly visible in the lens, while the blurred background suggests a late-night study. The entire picture is shrouded in an intense silence, creating an atmosphere of extreme concentration, tension and inner struggle, allowing the audience to directly feel the intense thinking going on in his mind.", + "seed": 26223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_11/index_5889_seed_26223_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030966224428266288, + "vlm_logprob": -0.030966224428266288, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.030966224428266288, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008465949445962906 + ], + "opposite_logprob_tokens": [ + -5.258465766906738 + ], + "logprob_reward": -0.008465949445962906, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008465949445962906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18669821321964264, + 0.1501065343618393, + 0.034733109176158905, + 0.06543755531311035, + 0.10400502383708954, + 0.025340735912322998, + 0.029880346730351448, + 0.07839953899383545, + 0.22488659620285034, + 0.1005123108625412 + ], + "credit_map": [ + 0.017129618674516678, + 0.02104239910840988, + 0.01647220365703106, + 0.020218614488840103, + 0.02441236563026905, + 0.03814805671572685, + 0.022555330768227577, + 0.02671964280307293, + 0.01258113607764244, + 0.008051466196775436, + 0.04526565596461296, + 0.007996194064617157, + 0.011174817569553852, + 0.013761120848357677, + 0.008548058569431305, + 0.04272808879613876, + 0.003418572712689638, + 0.0026152380742132664, + 0.003580836346372962, + 0.0037650729063898325, + 0.004732018336653709, + 0.006386230234056711, + 0.0039787390269339085, + 0.00625640107318759, + 0.005121149122714996, + 0.007246149703860283, + 0.016630131751298904, + 0.004919910803437233, + 0.0052526663057506084, + 0.007024122402071953, + 0.004979117773473263, + 0.014264310710132122, + 0.008946249261498451, + 0.007573303766548634, + 0.030001945793628693, + 0.007078797556459904, + 0.008165935985744, + 0.00654238136485219, + 0.007436380721628666, + 0.02826002798974514, + 0.0019425743957981467, + 0.002754182554781437, + 0.005600546952337027, + 0.001962176524102688, + 0.001985353184863925, + 0.0026152399368584156, + 0.002376079559326172, + 0.006104583851993084, + 0.0032264641486108303, + 0.0026091178879141808, + 0.006247494835406542, + 0.0022371243685483932, + 0.0028271444607526064, + 0.002556445775553584, + 0.003662907052785158, + 0.006513649597764015, + 0.007095866370946169, + 0.002683915663510561, + 0.026990050449967384, + 0.0026287336368113756, + 0.005352278705686331, + 0.004848184995353222, + 0.0029664565809071064, + 0.02583405002951622, + 0.020751675590872765, + 0.008915728889405727, + 0.0750705748796463, + 0.008029143325984478, + 0.015622830018401146, + 0.011738230474293232, + 0.010581274516880512, + 0.07417713850736618, + 0.007724130991846323, + 0.011148925870656967, + 0.023822328075766563, + 0.005130335222929716, + 0.008633404038846493, + 0.00716547342017293, + 0.006947518326342106, + 0.029940195381641388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.008465949445962906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020607998594641685 + ], + "opposite_logprob_tokens": [ + -3.8956079483032227 + ], + "logprob_reward": -0.020607998594641685, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020607998594641685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18682092428207397, + 0.07976984232664108, + 0.06370814889669418, + 0.10763852298259735, + 0.1335398107767105, + 0.0685134008526802, + 0.05417809635400772, + 0.08045731484889984, + 0.18204574286937714, + 0.04332815483212471 + ], + "credit_map": [ + 0.016423476859927177, + 0.01601455546915531, + 0.049636516720056534, + 0.017899665981531143, + 0.017378168180584908, + 0.015928229317069054, + 0.02239948883652687, + 0.031140828505158424, + 0.0072079068049788475, + 0.008213000372052193, + 0.019919365644454956, + 0.005586185958236456, + 0.005723231006413698, + 0.00894063152372837, + 0.006611119024455547, + 0.017568400129675865, + 0.00624864362180233, + 0.0104147894307971, + 0.009340728633105755, + 0.006627229042351246, + 0.008006688207387924, + 0.007972906343638897, + 0.007815157063305378, + 0.007282011676579714, + 0.013385523110628128, + 0.00822240486741066, + 0.03341604024171829, + 0.008211134932935238, + 0.009937022812664509, + 0.00815203133970499, + 0.009300974197685719, + 0.017013393342494965, + 0.011951910331845284, + 0.012293322011828423, + 0.04081973060965538, + 0.00928136520087719, + 0.00913339015096426, + 0.006301980931311846, + 0.009976486675441265, + 0.033781636506319046, + 0.005030701868236065, + 0.008729919791221619, + 0.014595243148505688, + 0.005739940330386162, + 0.005573490168899298, + 0.00589310098439455, + 0.0061866771429777145, + 0.016764327883720398, + 0.004387627821415663, + 0.008956700563430786, + 0.007209446281194687, + 0.005743382032960653, + 0.006006227806210518, + 0.005675225984305143, + 0.007176684215664864, + 0.009022803977131844, + 0.006282416172325611, + 0.008036787621676922, + 0.021945448592305183, + 0.007696317974478006, + 0.005962604656815529, + 0.005281638354063034, + 0.007456727791577578, + 0.017795374616980553, + 0.015802331268787384, + 0.009963912889361382, + 0.05700555443763733, + 0.007242630235850811, + 0.013266222551465034, + 0.010388851165771484, + 0.010157166048884392, + 0.05821909010410309, + 0.003729667281731963, + 0.007629053667187691, + 0.006855306215584278, + 0.003442547284066677, + 0.00471494859084487, + 0.0036688235122710466, + 0.004105325788259506, + 0.00918248388916254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.020607998594641685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102980252355337 + ], + "opposite_logprob_tokens": [ + -5.504103183746338 + ], + "logprob_reward": -0.004102980252355337, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102980252355337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11095262318849564, + 0.11864738166332245, + 0.09405945986509323, + 0.29421505331993103, + 0.14609014987945557, + 0.090471550822258, + 0.041930925101041794, + 0.026369908824563026, + 0.02798835001885891, + 0.04927452653646469 + ], + "credit_map": [ + 0.011236180551350117, + 0.011141743510961533, + 0.011097244918346405, + 0.013235957361757755, + 0.016155660152435303, + 0.01257102657109499, + 0.016369827091693878, + 0.019145000725984573, + 0.012234815396368504, + 0.010151728056371212, + 0.03513263165950775, + 0.008215160109102726, + 0.009539138525724411, + 0.010385164059698582, + 0.009223432280123234, + 0.023765334859490395, + 0.006296749692410231, + 0.010461363010108471, + 0.01480649970471859, + 0.009280518628656864, + 0.011447030119597912, + 0.009643588215112686, + 0.014749833382666111, + 0.017373893409967422, + 0.03499196842312813, + 0.021564945578575134, + 0.10014771670103073, + 0.020017066970467567, + 0.024189239367842674, + 0.01580716483294964, + 0.021221375092864037, + 0.056275613605976105, + 0.011680534109473228, + 0.029336247593164444, + 0.02158832550048828, + 0.01500672660768032, + 0.014194648712873459, + 0.014648224227130413, + 0.01708792708814144, + 0.022547533735632896, + 0.010640633292496204, + 0.010192984715104103, + 0.026941154152154922, + 0.006446455605328083, + 0.009299899451434612, + 0.006403615232557058, + 0.00694731529802084, + 0.013599508441984653, + 0.004753248766064644, + 0.006740013137459755, + 0.006496046669781208, + 0.0040590763092041016, + 0.003650150028988719, + 0.004147855564951897, + 0.005348856560885906, + 0.0067356848157942295, + 0.002893277909606695, + 0.003734201192855835, + 0.0031399501021951437, + 0.0027342969551682472, + 0.00323338620364666, + 0.003960875794291496, + 0.00355541636236012, + 0.0031185096595436335, + 0.0027230968698859215, + 0.004174964036792517, + 0.005658366251736879, + 0.0021625934168696404, + 0.0028398435097187757, + 0.002910452429205179, + 0.002918030135333538, + 0.004601007327437401, + 0.003918107133358717, + 0.007930943742394447, + 0.0072312806732952595, + 0.005032016895711422, + 0.0052420529536902905, + 0.004074382595717907, + 0.00575976399704814, + 0.010085983201861382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.004102980252355337 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648913353681564 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.008648913353681564, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648913353681564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3232375681400299, + 0.22525432705879211, + 0.06507445126771927, + 0.08829221874475479, + 0.09266199171543121, + 0.05529749393463135, + 0.027494248002767563, + 0.06080397590994835, + 0.027742736041545868, + 0.034141045063734055 + ], + "credit_map": [ + 0.028223661705851555, + 0.01845463737845421, + 0.10053599625825882, + 0.017238274216651917, + 0.024168068543076515, + 0.018270796164870262, + 0.02226637676358223, + 0.09407971054315567, + 0.021382907405495644, + 0.012625861912965775, + 0.08151071518659592, + 0.011090736836194992, + 0.015178845264017582, + 0.010523518547415733, + 0.011152313090860844, + 0.06178940087556839, + 0.00659811170771718, + 0.005625239107757807, + 0.01803029142320156, + 0.006690613459795713, + 0.005351061932742596, + 0.005180029198527336, + 0.0059585305862128735, + 0.011640569195151329, + 0.008982534520328045, + 0.014886491931974888, + 0.01715477742254734, + 0.01093765627592802, + 0.008646391332149506, + 0.008269638754427433, + 0.010972541756927967, + 0.008442184887826443, + 0.008028974756598473, + 0.005510467104613781, + 0.029419541358947754, + 0.004417997319251299, + 0.006878765299916267, + 0.005046714097261429, + 0.006215205881744623, + 0.02714431658387184, + 0.005702110938727856, + 0.0046342830173671246, + 0.017922958359122276, + 0.002668897621333599, + 0.004394650459289551, + 0.0039560203440487385, + 0.0038451324217021465, + 0.012173430994153023, + 0.0030319252982735634, + 0.0024090816732496023, + 0.007566551212221384, + 0.0018910830840468407, + 0.002070855814963579, + 0.0023346850648522377, + 0.0016005454817786813, + 0.006589517462998629, + 0.005811510607600212, + 0.00283765303902328, + 0.01814853958785534, + 0.0028578736819326878, + 0.004390131216496229, + 0.0038613146170973778, + 0.0032937258947640657, + 0.019603228196501732, + 0.0028169804718345404, + 0.005718343425542116, + 0.005799364298582077, + 0.0024881355930119753, + 0.00290058646351099, + 0.0022213070187717676, + 0.002781294984742999, + 0.0030167216900736094, + 0.0025910406839102507, + 0.00288344151340425, + 0.00770930852741003, + 0.0025670763570815325, + 0.0028200489468872547, + 0.0030238910112529993, + 0.0032288229558616877, + 0.009317411109805107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.008648913353681564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300528049468994 + ], + "opposite_logprob_tokens": [ + -2.2380051612854004 + ], + "logprob_reward": -0.11300528049468994, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300528049468994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04202602803707123, + 0.09950174391269684, + 0.050883498042821884, + 0.3205038607120514, + 0.24582993984222412, + 0.022914329543709755, + 0.057047031819820404, + 0.08587246388196945, + 0.039248671382665634, + 0.03617243096232414 + ], + "credit_map": [ + 0.0041365488432347775, + 0.005447165574878454, + 0.0062634628266096115, + 0.005126290954649448, + 0.004299672786146402, + 0.00494824443012476, + 0.005489674862474203, + 0.006314966827630997, + 0.011891362257301807, + 0.012133880518376827, + 0.02742086537182331, + 0.006872035562992096, + 0.009920161217451096, + 0.01146407425403595, + 0.006970085669308901, + 0.012829283252358437, + 0.005151406861841679, + 0.004067334346473217, + 0.01528141275048256, + 0.003056526416912675, + 0.004116833209991455, + 0.003328129881992936, + 0.004511197563260794, + 0.01137065701186657, + 0.03506092354655266, + 0.036518555134534836, + 0.09070247411727905, + 0.02874823659658432, + 0.02651405893266201, + 0.021957531571388245, + 0.0312928780913353, + 0.04970921203494072, + 0.023600062355399132, + 0.057125791907310486, + 0.03271418809890747, + 0.027541393414139748, + 0.022475849837064743, + 0.02095586247742176, + 0.030138997361063957, + 0.031277772039175034, + 0.00278886454179883, + 0.002778145484626293, + 0.0053125107660889626, + 0.0020043684635311365, + 0.001946918317116797, + 0.0023407230619341135, + 0.0031551416032016277, + 0.002587656257674098, + 0.005746297538280487, + 0.006319193169474602, + 0.013139665126800537, + 0.0032689126674085855, + 0.004744139965623617, + 0.0048751067370176315, + 0.004667660221457481, + 0.014286058023571968, + 0.007777313701808453, + 0.0029026889242231846, + 0.028830688446760178, + 0.003332706866785884, + 0.0058279503136873245, + 0.004827080760151148, + 0.0032646958716213703, + 0.02910934016108513, + 0.002687172731384635, + 0.004759257659316063, + 0.0053459834307432175, + 0.004215439781546593, + 0.004316153936088085, + 0.0045087337493896484, + 0.00519853550940752, + 0.008217391557991505, + 0.0035026038531214, + 0.007845727726817131, + 0.005400171969085932, + 0.0029997287783771753, + 0.0038517629727721214, + 0.002994789043441415, + 0.003918949048966169, + 0.005658697336912155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.11300528049468994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot captured from a level angle, focuses squarely on the face of a middle-aged man, his brow furrowed in deep thought over a vexing problem. A single, warm light source, like the light in a Rembrandt painting, dramatically sculpts his features, illuminating one cheek and making tiny beads of sweat glisten on his temple, while the other side is shrouded in deep shadow. Deep vertical wrinkles form between his brows from intense concentration, while horizontal lines of thought crisscross his forehead. His eyes are slightly narrowed, his gaze unfocused, as if gazing into the distance through the lens. His lips are pressed into a firm, resolute line, his jaw muscles subtly tensed, betraying an inner struggle. A few strands of dark, slightly damp hair fall across his forehead, the texture and pores of his skin clearly visible in the lens, while the blurred background suggests a late-night study. The entire picture is shrouded in an intense silence, creating an atmosphere of extreme concentration, tension and inner struggle, allowing the audience to directly feel the intense thinking going on in his mind.", + "seed": 26224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_11/index_5889_seed_26224_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030813723523169755, + "vlm_logprob": -0.030813723523169755, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.030813723523169755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006024297792464495 + ], + "opposite_logprob_tokens": [ + -5.881024360656738 + ], + "logprob_reward": -0.006024297792464495, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006024297792464495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.258037805557251, + 0.12852239608764648, + 0.19515854120254517, + 0.07152243703603745, + 0.06779950857162476, + 0.031597353518009186, + 0.08995919674634933, + 0.042509932070970535, + 0.043553344905376434, + 0.07133948802947998 + ], + "credit_map": [ + 0.04077976942062378, + 0.020245114341378212, + 0.03539448603987694, + 0.018289299681782722, + 0.022966433316469193, + 0.04566941037774086, + 0.037486642599105835, + 0.03720663860440254, + 0.027210719883441925, + 0.01909816265106201, + 0.015859341248869896, + 0.011649487540125847, + 0.010862279683351517, + 0.015566244721412659, + 0.014463636092841625, + 0.013812515884637833, + 0.02038932405412197, + 0.022580040618777275, + 0.02523105777800083, + 0.023081989958882332, + 0.017852703109383583, + 0.034612610936164856, + 0.026393091306090355, + 0.02501770667731762, + 0.009572966955602169, + 0.008996973745524883, + 0.009329862892627716, + 0.008214478380978107, + 0.008532705716788769, + 0.008429070934653282, + 0.012309123761951923, + 0.006137250456959009, + 0.00741592887789011, + 0.01318768598139286, + 0.00807725079357624, + 0.0075495317578315735, + 0.007184953428804874, + 0.01189768873155117, + 0.006442282348871231, + 0.006044177804142237, + 0.0035369866527616978, + 0.004491919185966253, + 0.0036781791131943464, + 0.0043270220048725605, + 0.0026273783296346664, + 0.003763071494176984, + 0.005384984891861677, + 0.0037878116127103567, + 0.009057636372745037, + 0.011649218387901783, + 0.014631014317274094, + 0.00774482311680913, + 0.00945739820599556, + 0.02174401469528675, + 0.007927984930574894, + 0.0077471062541007996, + 0.003615069203078747, + 0.004767217673361301, + 0.007328996900469065, + 0.0034411295782774687, + 0.005513329058885574, + 0.00728639168664813, + 0.006127712316811085, + 0.004430081229656935, + 0.0047788964584469795, + 0.0035308056976646185, + 0.006301072426140308, + 0.0036310749128460884, + 0.003276546485722065, + 0.01055319607257843, + 0.004718533251434565, + 0.0067632170394063, + 0.00953757669776678, + 0.008770747110247612, + 0.007492789067327976, + 0.007050972897559404, + 0.00785139761865139, + 0.009952006861567497, + 0.013573568314313889, + 0.007110426668077707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.006024297792464495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020606713369488716 + ], + "opposite_logprob_tokens": [ + -3.895606756210327 + ], + "logprob_reward": -0.020606713369488716, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020606713369488716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12716735899448395, + 0.04944867268204689, + 0.09409428387880325, + 0.10417192429304123, + 0.06128684803843498, + 0.0877186581492424, + 0.23673123121261597, + 0.10548845678567886, + 0.06627079844474792, + 0.06762178987264633 + ], + "credit_map": [ + 0.023376623168587685, + 0.02006014809012413, + 0.017343729734420776, + 0.01354079507291317, + 0.009524918161332607, + 0.016032079234719276, + 0.011217878200113773, + 0.016071181744337082, + 0.0054277279414236546, + 0.005481765139847994, + 0.005920999217778444, + 0.0067907581105828285, + 0.0051200189627707005, + 0.006437189411371946, + 0.009813605807721615, + 0.004456607159227133, + 0.009297280572354794, + 0.00594763970002532, + 0.018792590126395226, + 0.004643001593649387, + 0.006368208676576614, + 0.028890598565340042, + 0.0074512236751616, + 0.012703731656074524, + 0.011832989752292633, + 0.022999703884124756, + 0.013616354204714298, + 0.014092562720179558, + 0.011751282960176468, + 0.012715956196188927, + 0.01040918193757534, + 0.006753884255886078, + 0.007016919553279877, + 0.005842629354447126, + 0.007979667745530605, + 0.004154023248702288, + 0.006282277870923281, + 0.017298655584454536, + 0.005938696209341288, + 0.006773978006094694, + 0.013052785769104958, + 0.013266255147755146, + 0.010542495176196098, + 0.00860877986997366, + 0.007093470077961683, + 0.015014713630080223, + 0.009048499166965485, + 0.011091668158769608, + 0.015555892139673233, + 0.014323489740490913, + 0.047376543283462524, + 0.010256760753691196, + 0.019888190552592278, + 0.08104664087295532, + 0.01513686403632164, + 0.03314685449004173, + 0.008043564856052399, + 0.012638166546821594, + 0.011498240754008293, + 0.009988455101847649, + 0.011444327421486378, + 0.023291584104299545, + 0.015453113242983818, + 0.01313100103288889, + 0.005098766181617975, + 0.007722729817032814, + 0.006814069580286741, + 0.006021228153258562, + 0.007139924447983503, + 0.011550984345376492, + 0.012998639605939388, + 0.008924451656639576, + 0.006266285665333271, + 0.01310208160430193, + 0.007969171740114689, + 0.009917901828885078, + 0.006660847924649715, + 0.007913478650152683, + 0.00903070718050003, + 0.006761311553418636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.020606713369488716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046457452699542046 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.0046457452699542046, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046457452699542046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11156024783849716, + 0.1681230366230011, + 0.12522459030151367, + 0.0851132869720459, + 0.06339661777019501, + 0.1057581678032875, + 0.07420402020215988, + 0.06495228409767151, + 0.07527720928192139, + 0.12639057636260986 + ], + "credit_map": [ + 0.015511141158640385, + 0.016080453991889954, + 0.01743205450475216, + 0.01103999838232994, + 0.010826701298356056, + 0.016894733533263206, + 0.009441529400646687, + 0.014333619736135006, + 0.026548482477664948, + 0.01897667907178402, + 0.021166810765862465, + 0.016896814107894897, + 0.011080624535679817, + 0.03039611503481865, + 0.02025093138217926, + 0.022806573659181595, + 0.013509091921150684, + 0.011603212915360928, + 0.02126617357134819, + 0.011680783703923225, + 0.009717115201056004, + 0.029746603220701218, + 0.012930554337799549, + 0.0147710544988513, + 0.017178066074848175, + 0.010187342762947083, + 0.012090814299881458, + 0.00800384022295475, + 0.0075031453743577, + 0.010562736541032791, + 0.010516354814171791, + 0.009070982225239277, + 0.00923734437674284, + 0.00944635458290577, + 0.009628726169466972, + 0.006180175580084324, + 0.004714094568043947, + 0.009019752964377403, + 0.006508341990411282, + 0.008661825209856033, + 0.008060110732913017, + 0.009783972054719925, + 0.014305167831480503, + 0.005623405333608389, + 0.00959580298513174, + 0.028907397761940956, + 0.012971121817827225, + 0.016511185094714165, + 0.007241532206535339, + 0.011080902069807053, + 0.010183362290263176, + 0.00701048132032156, + 0.00800217967480421, + 0.016458414494991302, + 0.005878593306988478, + 0.008348548784852028, + 0.00521269254386425, + 0.0047255321405828, + 0.010568092577159405, + 0.003908204846084118, + 0.006505114492028952, + 0.019050609320402145, + 0.0065141404047608376, + 0.008467894978821278, + 0.008388128131628036, + 0.007778279948979616, + 0.011503293178975582, + 0.006096513010561466, + 0.006974419113248587, + 0.01721736416220665, + 0.00887820590287447, + 0.008440998382866383, + 0.01191084273159504, + 0.013788829557597637, + 0.021880947053432465, + 0.010643520392477512, + 0.009490903466939926, + 0.03255188465118408, + 0.012898210436105728, + 0.013225434347987175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.0046457452699542046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979179609566927 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.00979179609566927, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979179609566927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1321650892496109, + 0.06873515248298645, + 0.05376264825463295, + 0.14047127962112427, + 0.036762386560440063, + 0.0691043809056282, + 0.039974719285964966, + 0.11028239130973816, + 0.15311826765537262, + 0.19562365114688873 + ], + "credit_map": [ + 0.012642601504921913, + 0.017346573993563652, + 0.018835242837667465, + 0.014719734899699688, + 0.014733266085386276, + 0.01779092662036419, + 0.026611950248479843, + 0.009484808892011642, + 0.008334776386618614, + 0.007587487809360027, + 0.008480586111545563, + 0.005975988693535328, + 0.006235032808035612, + 0.011509086936712265, + 0.010521047748625278, + 0.010091148316860199, + 0.006727926433086395, + 0.004901431966573, + 0.00816053431481123, + 0.0037942007184028625, + 0.004825590644031763, + 0.012413754127919674, + 0.006203815806657076, + 0.006735401693731546, + 0.02768317610025406, + 0.016972938552498817, + 0.019777080044150352, + 0.013084295205771923, + 0.012926063500344753, + 0.015882618725299835, + 0.01827928051352501, + 0.01586584374308586, + 0.003047906793653965, + 0.005816191900521517, + 0.0036250220146030188, + 0.004060244653373957, + 0.003950048703700304, + 0.004056557081639767, + 0.00777037488296628, + 0.004436044953763485, + 0.011780863627791405, + 0.005711881909519434, + 0.00880634505301714, + 0.0050062392838299274, + 0.005840654019266367, + 0.012819190509617329, + 0.01101724710315466, + 0.00812196359038353, + 0.005412576254457235, + 0.00666050985455513, + 0.005917839240282774, + 0.004353804513812065, + 0.0035498831421136856, + 0.005544763524085283, + 0.004731946624815464, + 0.00380339496769011, + 0.009783897548913956, + 0.012760180979967117, + 0.014339298009872437, + 0.009072779677808285, + 0.01075268816202879, + 0.02717999368906021, + 0.01284441165626049, + 0.013549142517149448, + 0.012079989537596703, + 0.008140003308653831, + 0.019482459872961044, + 0.011744438670575619, + 0.015377072617411613, + 0.038314979523420334, + 0.027476316317915916, + 0.020503021776676178, + 0.021076487377285957, + 0.04213586077094078, + 0.02497880719602108, + 0.02579618990421295, + 0.017634758725762367, + 0.026108063757419586, + 0.019212929531931877, + 0.01868056133389473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.00979179609566927 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing squarely on the face of a middle-aged man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static, captured from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single warm light source illuminating one cheek while the other side is in deep shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible details such as deep vertical wrinkles between the brows, horizontal lines on the forehead, and tiny beads of sweat present?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred to suggest a late-night study environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1130000650882721 + ], + "opposite_logprob_tokens": [ + -2.2380001544952393 + ], + "logprob_reward": -0.1130000650882721, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1130000650882721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1375199556350708, + 0.11924312263727188, + 0.06257134675979614, + 0.07125359773635864, + 0.15797756612300873, + 0.13134479522705078, + 0.14702455699443817, + 0.05510801821947098, + 0.053784046322107315, + 0.06417298316955566 + ], + "credit_map": [ + 0.012994682416319847, + 0.018356965854763985, + 0.01834452711045742, + 0.01626593992114067, + 0.015153491869568825, + 0.019916972145438194, + 0.020480461418628693, + 0.016006913036108017, + 0.02234378457069397, + 0.013039723038673401, + 0.014241624623537064, + 0.012354876846075058, + 0.01169304084032774, + 0.01432061567902565, + 0.017507005482912064, + 0.013742458075284958, + 0.007706682663410902, + 0.013519114814698696, + 0.006805060897022486, + 0.00607025483623147, + 0.006226010154932737, + 0.007384677417576313, + 0.009181587025523186, + 0.005677967797964811, + 0.007274238858371973, + 0.012599149718880653, + 0.009639369323849678, + 0.00709467614069581, + 0.008261223323643208, + 0.009488477371633053, + 0.010618734173476696, + 0.0062777274288237095, + 0.01808459311723709, + 0.032080914825201035, + 0.01757953129708767, + 0.015567760914564133, + 0.015206092968583107, + 0.023444436490535736, + 0.0175199992954731, + 0.018494244664907455, + 0.025438686832785606, + 0.015871401876211166, + 0.016655046492815018, + 0.011376475915312767, + 0.011067464016377926, + 0.017156457528471947, + 0.02010047249495983, + 0.013678798452019691, + 0.014390534721314907, + 0.016082098707556725, + 0.0295315682888031, + 0.01084403321146965, + 0.011341852135956287, + 0.04057268798351288, + 0.010197395458817482, + 0.014064397662878036, + 0.0051545752212405205, + 0.009759106673300266, + 0.0066629694774746895, + 0.005904478952288628, + 0.004987979307770729, + 0.011214861646294594, + 0.004971770104020834, + 0.006452281028032303, + 0.003266387851908803, + 0.005347931757569313, + 0.008937599137425423, + 0.006172461900860071, + 0.003888914827257395, + 0.013891631737351418, + 0.006647367030382156, + 0.005631755106151104, + 0.0050197383388876915, + 0.015406192280352116, + 0.008069339208304882, + 0.009266200475394726, + 0.006498556584119797, + 0.00799508672207594, + 0.00633141677826643, + 0.005586457438766956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.1130000650882721 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_12/index_tensor([5224, 5224])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_12/index_tensor([5224, 5224])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..322391869605b3076d2b004e8a2d3386e31f0b31 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_12/index_tensor([5224, 5224])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly advancing close-up is fixed on the screen of an old cathode-ray tube (CRT) monitor. The monitor's thick, beige plastic case, slightly yellowed by age, is covered in tiny scratches and dust. In the center of its slightly convex, bottomless black glass screen, a stick figure, composed entirely of bright, slightly glowing phosphorescent green vector lines, steps backward in a mechanical, clumsy manner. Its movements are not fluid animations, but rather a slow, staccato flow: one leg stiffly raises and lowers backward, and the other follows in an equally uncoordinated manner, its entire body flickering faintly and jumping between frames. This green figure exists isolated in the pure blackness of digital space, its cold glow the sole primary source of light, casting a retro and slightly mysterious aura upon its gloomy surroundings, creating a strange sense of contemplation upon the lonely operation of outdated technology.", + "seed": 7989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_12/index_5224_seed_7989_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021575834904797375, + "vlm_logprob": -0.021575834904797375, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.021575834904797375, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005337986163794994 + ], + "opposite_logprob_tokens": [ + -6.380338191986084 + ], + "logprob_reward": -0.005337986163794994, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.005337986163794994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26844239234924316, + 0.1838941127061844, + 0.04208127036690712, + 0.05142451077699661, + 0.15094231069087982, + 0.02237444743514061, + 0.12747915089130402, + 0.018038401380181313, + 0.09137815237045288, + 0.04394528642296791 + ], + "credit_map": [ + 0.02286420203745365, + 0.026113370433449745, + 0.0859292671084404, + 0.03011227212846279, + 0.025347910821437836, + 0.0205899178981781, + 0.027173243463039398, + 0.030312195420265198, + 0.01708086207509041, + 0.009388416074216366, + 0.06666452437639236, + 0.029568316414952278, + 0.008053401485085487, + 0.017530575394630432, + 0.00864794384688139, + 0.026960065588355064, + 0.003966980148106813, + 0.007456777151674032, + 0.011309958063066006, + 0.0038910445291548967, + 0.004113988950848579, + 0.0029694365803152323, + 0.0033635124564170837, + 0.00500957015901804, + 0.004754568450152874, + 0.004888107068836689, + 0.014598779380321503, + 0.006071840412914753, + 0.004037571605294943, + 0.003856202820315957, + 0.006722918711602688, + 0.006494523957371712, + 0.009073321707546711, + 0.020904995501041412, + 0.035261571407318115, + 0.046983398497104645, + 0.005832029972225428, + 0.010047303512692451, + 0.013732423074543476, + 0.009107263758778572, + 0.0019099742639809847, + 0.005924717988818884, + 0.0011810562573373318, + 0.0021783087868243456, + 0.004093995783478022, + 0.002056376775726676, + 0.002665008883923292, + 0.0023650091607123613, + 0.011570712551474571, + 0.010694026947021484, + 0.04024453088641167, + 0.015310387127101421, + 0.013497925363481045, + 0.009904933162033558, + 0.00813357438892126, + 0.018123064190149307, + 0.0016205663559958339, + 0.0022174750920385122, + 0.004034535959362984, + 0.00207823165692389, + 0.0021663911174982786, + 0.0019455691799521446, + 0.002500419272109866, + 0.0014752125134691596, + 0.007896357215940952, + 0.020790962502360344, + 0.02489549294114113, + 0.009286019951105118, + 0.00987234991043806, + 0.005898430477827787, + 0.006492475979030132, + 0.006246059667319059, + 0.003944205120205879, + 0.003397207008674741, + 0.015432560816407204, + 0.005148998461663723, + 0.003401718335226178, + 0.0038808751851320267, + 0.0025997913908213377, + 0.0061399261467158794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005337986163794994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011103389784693718 + ], + "opposite_logprob_tokens": [ + -4.51110315322876 + ], + "logprob_reward": -0.011103389784693718, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011103389784693718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1586848646402359, + 0.18683505058288574, + 0.019763918593525887, + 0.021731095388531685, + 0.30614566802978516, + 0.015945930033922195, + 0.018349846825003624, + 0.00927706342190504, + 0.20924808084964752, + 0.05401849001646042 + ], + "credit_map": [ + 0.01803744025528431, + 0.010214096866548061, + 0.05882745236158371, + 0.017906058579683304, + 0.00865246169269085, + 0.01784081943333149, + 0.008956046774983406, + 0.01825050264596939, + 0.018020370975136757, + 0.013746264390647411, + 0.0700848177075386, + 0.026636021211743355, + 0.008276853710412979, + 0.016139604151248932, + 0.008192888461053371, + 0.02573823556303978, + 0.0019432362169027328, + 0.004036732017993927, + 0.0032728638034313917, + 0.00202268548309803, + 0.00242214766331017, + 0.0025500075425952673, + 0.0016940219793468714, + 0.0018222251674160361, + 0.0014075561193749309, + 0.003534110728651285, + 0.0028453231789171696, + 0.002680874662473798, + 0.003408607328310609, + 0.002072150819003582, + 0.004174146335572004, + 0.0016083266818895936, + 0.020346131175756454, + 0.0286243986338377, + 0.08251165598630905, + 0.09215309470891953, + 0.010116498917341232, + 0.022282881662249565, + 0.0289352685213089, + 0.021175751462578773, + 0.001614223001524806, + 0.0027802421245723963, + 0.001146528054960072, + 0.0015860686544328928, + 0.00232876418158412, + 0.0016064165392890573, + 0.002541369292885065, + 0.0023423200473189354, + 0.0009804802248254418, + 0.0022554288152605295, + 0.0007863802602514625, + 0.0031616308260709047, + 0.003160659922286868, + 0.0016574697801843286, + 0.004913982469588518, + 0.001433814875781536, + 0.000698232848662883, + 0.0009282963001169264, + 0.00151908490806818, + 0.0018765233689919114, + 0.0010986682027578354, + 0.0007912350702099502, + 0.001742980326525867, + 0.0006220431532710791, + 0.01730620115995407, + 0.04976913332939148, + 0.061105795204639435, + 0.019872907549142838, + 0.02127773128449917, + 0.01569836027920246, + 0.010746506042778492, + 0.013471449725329876, + 0.004163045901805162, + 0.005334567278623581, + 0.014020493254065514, + 0.005616916809231043, + 0.008382891304790974, + 0.0036547782365232706, + 0.005020020063966513, + 0.007825782522559166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.011103389784693718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008225633064284921 + ], + "opposite_logprob_tokens": [ + -7.1258225440979 + ], + "logprob_reward": -0.0008225633064284921, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008225633064284921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1814400851726532, + 0.15323182940483093, + 0.05079453065991402, + 0.09732942283153534, + 0.28200265765190125, + 0.03977097198367119, + 0.03627457469701767, + 0.05956731736660004, + 0.07138566672801971, + 0.0282029677182436 + ], + "credit_map": [ + 0.02528897300362587, + 0.01888369396328926, + 0.04952321574091911, + 0.016538776457309723, + 0.01513404119759798, + 0.020355883985757828, + 0.01834993064403534, + 0.01736557111144066, + 0.017613613978028297, + 0.0181878674775362, + 0.034268569201231, + 0.015651846304535866, + 0.01847452111542225, + 0.012408166192471981, + 0.02417278289794922, + 0.012454460375010967, + 0.004065832123160362, + 0.008559566922485828, + 0.003133014077320695, + 0.004985445644706488, + 0.012434353120625019, + 0.00513823376968503, + 0.007211078889667988, + 0.0052670082077383995, + 0.007234867662191391, + 0.012439564801752567, + 0.012440940365195274, + 0.011024109087884426, + 0.01274801790714264, + 0.00955866277217865, + 0.01716221496462822, + 0.014721041545271873, + 0.015624094754457474, + 0.05733342841267586, + 0.03383274003863335, + 0.09506481885910034, + 0.017652323469519615, + 0.014935293234884739, + 0.03424001857638359, + 0.013319933786988258, + 0.002679306548088789, + 0.008866195566952229, + 0.004207768011838198, + 0.004578791558742523, + 0.006397130899131298, + 0.004170748870819807, + 0.005627066362649202, + 0.003243966493755579, + 0.0025513579603284597, + 0.0039973510429263115, + 0.005335513968020678, + 0.004860871005803347, + 0.0043875244446098804, + 0.004125638399273157, + 0.007265646941959858, + 0.003750670701265335, + 0.004118870943784714, + 0.01536897849291563, + 0.011546377092599869, + 0.007043341174721718, + 0.007492775563150644, + 0.004109115805476904, + 0.006043922621756792, + 0.0038439342752099037, + 0.00629655085504055, + 0.012380794622004032, + 0.020463209599256516, + 0.007463244255632162, + 0.008598998188972473, + 0.00548201147466898, + 0.005554747302085161, + 0.0051461150869727135, + 0.0025070440024137497, + 0.004541151225566864, + 0.004546881187707186, + 0.0028920257464051247, + 0.005335566122084856, + 0.0021485870238393545, + 0.0040830946527421474, + 0.0021486179903149605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0008225633064284921 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003526389598846 + ], + "opposite_logprob_tokens": [ + -2.695035219192505 + ], + "logprob_reward": -0.07003526389598846, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003526389598846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2819865047931671, + 0.15037278831005096, + 0.034370988607406616, + 0.07323061674833298, + 0.19222207367420197, + 0.03706521913409233, + 0.03497316315770149, + 0.03340169042348862, + 0.1334010362625122, + 0.028975825756788254 + ], + "credit_map": [ + 0.028579968959093094, + 0.018744556233286858, + 0.10208054631948471, + 0.03127424791455269, + 0.017038291320204735, + 0.028211921453475952, + 0.020465778186917305, + 0.03559126704931259, + 0.012503579258918762, + 0.02285194583237171, + 0.02069893665611744, + 0.031698815524578094, + 0.018213968724012375, + 0.010101285763084888, + 0.019373951479792595, + 0.014930318109691143, + 0.003676407039165497, + 0.005676832515746355, + 0.0033108496572822332, + 0.003710865043103695, + 0.006642443593591452, + 0.0033528937492519617, + 0.00446453457698226, + 0.0035361687187105417, + 0.0058071063831448555, + 0.0068399375304579735, + 0.011369174346327782, + 0.008019272238016129, + 0.01010833214968443, + 0.007125799544155598, + 0.012779837474226952, + 0.011181163601577282, + 0.010647769086062908, + 0.03796911612153053, + 0.024038366973400116, + 0.06278639286756516, + 0.013142951764166355, + 0.009954330511391163, + 0.02357025071978569, + 0.01011289469897747, + 0.0025643890257924795, + 0.009153078310191631, + 0.0021867132745683193, + 0.00489892577752471, + 0.006098773330450058, + 0.0036311864387243986, + 0.005326868500560522, + 0.003205288900062442, + 0.002156478352844715, + 0.004019355401396751, + 0.005184152629226446, + 0.005213051103055477, + 0.003930170554667711, + 0.003214424941688776, + 0.008125725202262402, + 0.0031298103276640177, + 0.0025682509876787663, + 0.009362098760902882, + 0.005386802367866039, + 0.004636060446500778, + 0.004092190880328417, + 0.002287116600200534, + 0.002688419772312045, + 0.0023807547986507416, + 0.01153893768787384, + 0.031025731936097145, + 0.03937884792685509, + 0.01281216461211443, + 0.013860694132745266, + 0.010358201339840889, + 0.005446148104965687, + 0.008980323560535908, + 0.002097638323903084, + 0.003645122516900301, + 0.0031445061322301626, + 0.004178955685347319, + 0.0045478432439267635, + 0.002613662974908948, + 0.005942147225141525, + 0.002805953146889806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.07003526389598846 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579971373081207 + ], + "opposite_logprob_tokens": [ + -3.8955800533294678 + ], + "logprob_reward": -0.020579971373081207, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579971373081207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14649619162082672, + 0.21473023295402527, + 0.034002725034952164, + 0.08440669625997543, + 0.23355655372142792, + 0.04152370244264603, + 0.03453373908996582, + 0.06501415371894836, + 0.12537315487861633, + 0.020362915471196175 + ], + "credit_map": [ + 0.02111920155584812, + 0.013382133096456528, + 0.04206378757953644, + 0.013045167550444603, + 0.015138315968215466, + 0.016286849975585938, + 0.012176754884421825, + 0.013283967971801758, + 0.0206051766872406, + 0.032365329563617706, + 0.06697335839271545, + 0.023202748969197273, + 0.017752571031451225, + 0.014692883007228374, + 0.012101433239877224, + 0.027036722749471664, + 0.003813150804489851, + 0.00772102689370513, + 0.0030795421916991472, + 0.003434718120843172, + 0.005826547741889954, + 0.003344099037349224, + 0.0038363721687346697, + 0.0029472694732248783, + 0.007946120575070381, + 0.007586384192109108, + 0.01612190157175064, + 0.008333349600434303, + 0.008464143611490726, + 0.009081866592168808, + 0.011634773574769497, + 0.015238159336149693, + 0.014782085083425045, + 0.06133180111646652, + 0.008493845351040363, + 0.08244655281305313, + 0.01808302477002144, + 0.008752134628593922, + 0.029042182490229607, + 0.010624924674630165, + 0.0031987528782337904, + 0.009161152876913548, + 0.003336906898766756, + 0.0054840934462845325, + 0.006093671545386314, + 0.0044973548501729965, + 0.0057284291833639145, + 0.004023341462016106, + 0.002755506429821253, + 0.004105282947421074, + 0.006503701210021973, + 0.004703272134065628, + 0.003978054504841566, + 0.0031659246888011694, + 0.006528082769364119, + 0.002793913008645177, + 0.004439191427081823, + 0.015979204326868057, + 0.007515698205679655, + 0.011276565492153168, + 0.008384348824620247, + 0.004901626612991095, + 0.006023747380822897, + 0.006493767723441124, + 0.010168798267841339, + 0.02596435695886612, + 0.037743572145700455, + 0.012483098544180393, + 0.012240927666425705, + 0.008228201419115067, + 0.008792621083557606, + 0.009751594625413418, + 0.0019345205510035157, + 0.0026726725045591593, + 0.0026529552415013313, + 0.002606661058962345, + 0.003253818489611149, + 0.0020699722226709127, + 0.0034089621622115374, + 0.0017633531242609024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.020579971373081207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing close-up is fixed on the screen of an old cathode-ray tube (CRT) monitor. The monitor's thick, beige plastic case, slightly yellowed by age, is covered in tiny scratches and dust. In the center of its slightly convex, bottomless black glass screen, a stick figure, composed entirely of bright, slightly glowing phosphorescent green vector lines, steps backward in a mechanical, clumsy manner. Its movements are not fluid animations, but rather a slow, staccato flow: one leg stiffly raises and lowers backward, and the other follows in an equally uncoordinated manner, its entire body flickering faintly and jumping between frames. This green figure exists isolated in the pure blackness of digital space, its cold glow the sole primary source of light, casting a retro and slightly mysterious aura upon its gloomy surroundings, creating a strange sense of contemplation upon the lonely operation of outdated technology.", + "seed": 7990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_12/index_5224_seed_7990_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.034247824375052006, + "vlm_logprob": -0.034247824375052006, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.034247824375052006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022301126271486282 + ], + "opposite_logprob_tokens": [ + -4.022301197052002 + ], + "logprob_reward": -0.022301126271486282, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.022301126271486282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16752035915851593, + 0.029099686071276665, + 0.04480716958642006, + 0.11920730024576187, + 0.04289956018328667, + 0.42548736929893494, + 0.012360583990812302, + 0.1165209412574768, + 0.018840312957763672, + 0.023256629705429077 + ], + "credit_map": [ + 0.02658730000257492, + 0.010015582665801048, + 0.050421033054590225, + 0.016481615602970123, + 0.008222168311476707, + 0.010276669636368752, + 0.025567246600985527, + 0.019948752596974373, + 0.0043007992208004, + 0.002885108580812812, + 0.003315070178359747, + 0.004863849841058254, + 0.0034862158354371786, + 0.0013884356012567878, + 0.006222047843039036, + 0.002638160949572921, + 0.005401616916060448, + 0.002481826115399599, + 0.01527333538979292, + 0.004725972656160593, + 0.002721665194258094, + 0.0015996332513168454, + 0.006758444011211395, + 0.005844673607498407, + 0.01795065402984619, + 0.0042680599726736546, + 0.0447433739900589, + 0.012846347875893116, + 0.003864576341584325, + 0.0048132603988051414, + 0.013463707640767097, + 0.017257316038012505, + 0.006387131288647652, + 0.004045992158353329, + 0.00497771892696619, + 0.0055458033457398415, + 0.003883698023855686, + 0.0033648996613919735, + 0.006934720557183027, + 0.00775959761813283, + 0.04925895482301712, + 0.01949469931423664, + 0.15176424384117126, + 0.04104907065629959, + 0.012685262598097324, + 0.023966003209352493, + 0.03880208358168602, + 0.0884670615196228, + 0.0014873204054310918, + 0.0012745304265990853, + 0.0013169808080419898, + 0.0022662272676825523, + 0.001750511466525495, + 0.0007979319198057055, + 0.0024030657950788736, + 0.001064016600139439, + 0.02057027257978916, + 0.006194840185344219, + 0.03771904855966568, + 0.011659105308353901, + 0.0040150051936507225, + 0.006850046571344137, + 0.012330382131040096, + 0.017182249575853348, + 0.0024985179770737886, + 0.0014455465134233236, + 0.004585220944136381, + 0.0020304133649915457, + 0.0018327217549085617, + 0.0012084245681762695, + 0.002441015560179949, + 0.002798452042043209, + 0.0027812463231384754, + 0.0015132438857108355, + 0.00640365295112133, + 0.002982023637741804, + 0.002045875880867243, + 0.0014549465849995613, + 0.0032677664421498775, + 0.0028078739996999502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.022301126271486282 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08913596719503403 + ], + "opposite_logprob_tokens": [ + -2.4641358852386475 + ], + "logprob_reward": -0.08913596719503403, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08913596719503403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17046965658664703, + 0.02187694050371647, + 0.007719073444604874, + 0.0737311989068985, + 0.04813238978385925, + 0.5012973546981812, + 0.02759544737637043, + 0.1391277313232422, + 0.003970238380134106, + 0.00607998576015234 + ], + "credit_map": [ + 0.02499476447701454, + 0.008125930093228817, + 0.06146104633808136, + 0.015312033705413342, + 0.0057493532076478004, + 0.0072107925079762936, + 0.025007380172610283, + 0.022608356550335884, + 0.0031237362418323755, + 0.0018116910941898823, + 0.004860076121985912, + 0.002997973933815956, + 0.002025583293288946, + 0.0012020207941532135, + 0.0035427738912403584, + 0.002313085598871112, + 0.0008156709955073893, + 0.0007173221092671156, + 0.0010395020944997668, + 0.0012859510025009513, + 0.0008681475883349776, + 0.0006101407343521714, + 0.0017385749379172921, + 0.0006437641568481922, + 0.010732779279351234, + 0.0036047513131052256, + 0.026466868817806244, + 0.008378498256206512, + 0.0034186283592134714, + 0.003272969974204898, + 0.008442776277661324, + 0.009413926862180233, + 0.007108944933861494, + 0.004595158156007528, + 0.005274089053273201, + 0.006075523793697357, + 0.004383997991681099, + 0.0032811053097248077, + 0.00796232558786869, + 0.00945124588906765, + 0.0584292896091938, + 0.022958889603614807, + 0.1806228905916214, + 0.048132047057151794, + 0.014537576586008072, + 0.02740476280450821, + 0.0463608019053936, + 0.10285107046365738, + 0.0031163275707513094, + 0.0012900667497888207, + 0.010232537984848022, + 0.002538140630349517, + 0.0013502364745363593, + 0.0012567220255732536, + 0.0029626304749399424, + 0.004848785232752562, + 0.022017884999513626, + 0.008726473897695541, + 0.043430328369140625, + 0.015190942212939262, + 0.006402456667274237, + 0.006909646559506655, + 0.01714150793850422, + 0.01930849812924862, + 0.0005972206126898527, + 0.00029394146986305714, + 0.0010361672611907125, + 0.0004619751125574112, + 0.0003367826866451651, + 0.00030729410354979336, + 0.000533998419996351, + 0.00040285891736857593, + 0.0008006177376955748, + 0.0007256586104631424, + 0.0004704749444499612, + 0.0008926589507609606, + 0.0007373365806415677, + 0.0005996871041134, + 0.0011399624636396766, + 0.0007135892519727349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.08913596719503403 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006529344827868044 + ], + "opposite_logprob_tokens": [ + -7.37565279006958 + ], + "logprob_reward": -0.0006529344827868044, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006529344827868044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3957878351211548, + 0.05521460622549057, + 0.10687742382287979, + 0.049530029296875, + 0.026743831112980843, + 0.07993294298648834, + 0.013127697631716728, + 0.13299770653247833, + 0.07070183753967285, + 0.06908613443374634 + ], + "credit_map": [ + 0.05470694974064827, + 0.01724206656217575, + 0.14309273660182953, + 0.03461310639977455, + 0.011720564216375351, + 0.019205326214432716, + 0.0505254827439785, + 0.06468156725168228, + 0.0070954980328679085, + 0.004841769579797983, + 0.012126470915973186, + 0.007754243444651365, + 0.006482441443949938, + 0.002459547482430935, + 0.009823501110076904, + 0.004631131887435913, + 0.011724612675607204, + 0.00413468899205327, + 0.03597141057252884, + 0.010980530641973019, + 0.004732979927212, + 0.006440985016524792, + 0.0122977951541543, + 0.020594412460923195, + 0.007915966212749481, + 0.0034457340370863676, + 0.01111584436148405, + 0.0061012445949018, + 0.004520227666944265, + 0.003972665406763554, + 0.008375548757612705, + 0.004082799423485994, + 0.002661918057128787, + 0.002648959867656231, + 0.006341314874589443, + 0.0032623889856040478, + 0.002398944925516844, + 0.001484503853134811, + 0.004381466191262007, + 0.003564335172995925, + 0.010315768420696259, + 0.005752491764724255, + 0.023560555651783943, + 0.00989168044179678, + 0.00600383011624217, + 0.002791850594803691, + 0.012051383964717388, + 0.00956538412719965, + 0.0018527940846979618, + 0.0012428199406713247, + 0.002757427981123328, + 0.001723510678857565, + 0.0013627244625240564, + 0.0009036054252646863, + 0.0020974755752831697, + 0.001187338843010366, + 0.01575089991092682, + 0.011259380728006363, + 0.0360468253493309, + 0.018029311671853065, + 0.01061200350522995, + 0.004942985717207193, + 0.022730251774191856, + 0.013626044616103172, + 0.00984863005578518, + 0.00341209233738482, + 0.02470652386546135, + 0.006653068587183952, + 0.0033133237157016993, + 0.0029834937304258347, + 0.009224786423146725, + 0.010559916496276855, + 0.006755459122359753, + 0.004041413310915232, + 0.01786717027425766, + 0.00986082386225462, + 0.006266701500862837, + 0.0031593074090778828, + 0.01264570839703083, + 0.008489546366035938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0006529344827868044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043077558279037476 + ], + "opposite_logprob_tokens": [ + -3.1680774688720703 + ], + "logprob_reward": -0.043077558279037476, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043077558279037476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10245048999786377, + 0.051878344267606735, + 0.06432390958070755, + 0.0972069799900055, + 0.05125202611088753, + 0.20461592078208923, + 0.057442087680101395, + 0.3263668715953827, + 0.028254345059394836, + 0.016209037974476814 + ], + "credit_map": [ + 0.019484050571918488, + 0.006921567488461733, + 0.021172896027565002, + 0.009962701238691807, + 0.006913354154676199, + 0.00913702417165041, + 0.01479240134358406, + 0.014066499657928944, + 0.005915317218750715, + 0.004940115846693516, + 0.00727496575564146, + 0.008224370889365673, + 0.006064009387046099, + 0.0030958731658756733, + 0.010477757081389427, + 0.005885941907763481, + 0.008295693434774876, + 0.003299240255728364, + 0.019774913787841797, + 0.0077401455491781235, + 0.004929621238261461, + 0.0024527141358703375, + 0.010968071408569813, + 0.006863512098789215, + 0.011292250826954842, + 0.007179277949035168, + 0.02402162365615368, + 0.013820830732584, + 0.00810232199728489, + 0.0036716763861477375, + 0.02107047103345394, + 0.008048533461987972, + 0.005552268587052822, + 0.003232002956792712, + 0.013930125162005424, + 0.004794640466570854, + 0.004173429682850838, + 0.002919695805758238, + 0.006511176470667124, + 0.01013869047164917, + 0.019760290160775185, + 0.011991942301392555, + 0.06085899472236633, + 0.021534590050578117, + 0.010862186551094055, + 0.012203442864120007, + 0.026207435876131058, + 0.041197050362825394, + 0.006432025693356991, + 0.002600854728370905, + 0.02112962305545807, + 0.004969504661858082, + 0.0025678579695522785, + 0.002782425843179226, + 0.006234115920960903, + 0.010725686326622963, + 0.039161380380392075, + 0.02717399410903454, + 0.09915310889482498, + 0.04233914613723755, + 0.020709462463855743, + 0.011240162886679173, + 0.051252178847789764, + 0.03533743694424629, + 0.0036619151942431927, + 0.0018696993356570601, + 0.008631888777017593, + 0.003204586450010538, + 0.0016966897528618574, + 0.0018458609702065587, + 0.003705268260091543, + 0.00363843678496778, + 0.0017084443243220448, + 0.0015059858560562134, + 0.0019211292965337634, + 0.002459155162796378, + 0.0014732124982401729, + 0.0016931354766711593, + 0.0023276829160749912, + 0.0031202933751046658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.043077558279037476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071535646915436 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071535646915436, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071535646915436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3401907980442047, + 0.07148414105176926, + 0.058758314698934555, + 0.0482202023267746, + 0.033921126276254654, + 0.07376887649297714, + 0.052704211324453354, + 0.11457493901252747, + 0.033751484006643295, + 0.17262588441371918 + ], + "credit_map": [ + 0.0442146360874176, + 0.017918147146701813, + 0.11703314632177353, + 0.029601961374282837, + 0.014243016019463539, + 0.019102483987808228, + 0.04107150807976723, + 0.05700591579079628, + 0.008324665948748589, + 0.00759693281725049, + 0.00860907044261694, + 0.011984428390860558, + 0.007701749447733164, + 0.004089381080120802, + 0.01299047376960516, + 0.010187448933720589, + 0.00644865445792675, + 0.0038591634947806597, + 0.014795086346566677, + 0.007939960807561874, + 0.004886992275714874, + 0.0030843233689665794, + 0.011265737004578114, + 0.0064784022979438305, + 0.007632996421307325, + 0.005104450508952141, + 0.006409894209355116, + 0.006595679558813572, + 0.004159559495747089, + 0.0033843633718788624, + 0.009071123786270618, + 0.005862139165401459, + 0.0037811126094311476, + 0.0026795677840709686, + 0.004492396488785744, + 0.00477759912610054, + 0.00483575789257884, + 0.003691695397719741, + 0.005964502226561308, + 0.0036984963808208704, + 0.007793010678142309, + 0.00840017944574356, + 0.006390818860381842, + 0.014797505922615528, + 0.010052986443042755, + 0.003772337920963764, + 0.016842273995280266, + 0.005719763692468405, + 0.005618436262011528, + 0.003436170518398285, + 0.015567227266728878, + 0.00622017215937376, + 0.003599487477913499, + 0.0029557773377746344, + 0.007520727813243866, + 0.007786216679960489, + 0.009169905446469784, + 0.011896069161593914, + 0.021052425727248192, + 0.015820728614926338, + 0.012838730588555336, + 0.004957564175128937, + 0.023332154378294945, + 0.015507365576922894, + 0.0040654572658240795, + 0.0026502448599785566, + 0.008918301202356815, + 0.00420110858976841, + 0.0026775384321808815, + 0.002031614538282156, + 0.005934556480497122, + 0.0032726661302149296, + 0.023269694298505783, + 0.007390091195702553, + 0.061154939234256744, + 0.01664038375020027, + 0.005139448679983616, + 0.009234040975570679, + 0.018991347402334213, + 0.03080594912171364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.016071535646915436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing close-up is fixed on the screen of an old cathode-ray tube (CRT) monitor. The monitor's thick, beige plastic case, slightly yellowed by age, is covered in tiny scratches and dust. In the center of its slightly convex, bottomless black glass screen, a stick figure, composed entirely of bright, slightly glowing phosphorescent green vector lines, steps backward in a mechanical, clumsy manner. Its movements are not fluid animations, but rather a slow, staccato flow: one leg stiffly raises and lowers backward, and the other follows in an equally uncoordinated manner, its entire body flickering faintly and jumping between frames. This green figure exists isolated in the pure blackness of digital space, its cold glow the sole primary source of light, casting a retro and slightly mysterious aura upon its gloomy surroundings, creating a strange sense of contemplation upon the lonely operation of outdated technology.", + "seed": 7991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_12/index_5224_seed_7991_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03473796498728916, + "vlm_logprob": -0.03473796498728916, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.03473796498728916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011227876879274845 + ], + "opposite_logprob_tokens": [ + -5.261228084564209 + ], + "logprob_reward": -0.011227876879274845, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.011227876879274845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2899986505508423, + 0.1293586939573288, + 0.04540574550628662, + 0.06336534023284912, + 0.13427482545375824, + 0.07617897540330887, + 0.08349788933992386, + 0.07703644037246704, + 0.058746837079524994, + 0.042136598378419876 + ], + "credit_map": [ + 0.022343577817082405, + 0.06505445390939713, + 0.030332880094647408, + 0.03372295945882797, + 0.03290366008877754, + 0.050426702946424484, + 0.018690941855311394, + 0.0365234911441803, + 0.012245121411979198, + 0.02487899921834469, + 0.012770150788128376, + 0.01686589978635311, + 0.015749333426356316, + 0.016355637460947037, + 0.012839836068451405, + 0.017653722316026688, + 0.004978408105671406, + 0.007699146401137114, + 0.006029701791703701, + 0.005147653631865978, + 0.004731001798063517, + 0.007309035863727331, + 0.004221943207085133, + 0.005288855638355017, + 0.010155089199543, + 0.008133966475725174, + 0.008989390917122364, + 0.00541843194514513, + 0.00581869063898921, + 0.009895139373838902, + 0.005634126253426075, + 0.009320501238107681, + 0.009645526297390461, + 0.017329510301351547, + 0.022092506289482117, + 0.017311012372374535, + 0.01220876444131136, + 0.022317782044410706, + 0.013406677171587944, + 0.01996305026113987, + 0.00575345428660512, + 0.012816275469958782, + 0.004008653573691845, + 0.010811569169163704, + 0.0050391629338264465, + 0.014865800738334656, + 0.00587516138330102, + 0.01700889691710472, + 0.010778986848890781, + 0.009532628580927849, + 0.00999272521585226, + 0.007827412337064743, + 0.012561384588479996, + 0.013332155533134937, + 0.00932776927947998, + 0.010144824162125587, + 0.008551454171538353, + 0.009993929415941238, + 0.010615337640047073, + 0.007660472299903631, + 0.007016954477876425, + 0.012469429522752762, + 0.008444705046713352, + 0.012284159660339355, + 0.005702455062419176, + 0.010945496149361134, + 0.008967778645455837, + 0.009718555957078934, + 0.005402355454862118, + 0.006176547147333622, + 0.0054682157933712006, + 0.006365431938320398, + 0.004850319121032953, + 0.007329009938985109, + 0.004583803936839104, + 0.005518150981515646, + 0.004754398949444294, + 0.006611662916839123, + 0.0034948333632200956, + 0.004994419869035482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.011227876879274845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04312288761138916 + ], + "opposite_logprob_tokens": [ + -3.1681227684020996 + ], + "logprob_reward": -0.04312288761138916, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04312288761138916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2168845683336258, + 0.1260206401348114, + 0.06413553655147552, + 0.07761197537183762, + 0.21087546646595, + 0.07676514238119125, + 0.10252972692251205, + 0.06713055819272995, + 0.024238940328359604, + 0.03380745276808739 + ], + "credit_map": [ + 0.015348202548921108, + 0.04480034112930298, + 0.02210070937871933, + 0.044968798756599426, + 0.02141583524644375, + 0.026484068483114243, + 0.016813935711979866, + 0.024952661246061325, + 0.012689363211393356, + 0.02147234044969082, + 0.014453352428972721, + 0.014385289512574673, + 0.01607424020767212, + 0.01736588217318058, + 0.012595162726938725, + 0.016985006630420685, + 0.005657009314745665, + 0.009913752786815166, + 0.008278041146695614, + 0.007819157093763351, + 0.01035420224070549, + 0.007237776182591915, + 0.007354268338531256, + 0.007521322928369045, + 0.006636281497776508, + 0.010155865922570229, + 0.010603109374642372, + 0.010903170332312584, + 0.010329821147024632, + 0.011835237964987755, + 0.010034102015197277, + 0.0071143899112939835, + 0.019327417016029358, + 0.031190460547804832, + 0.029102103784680367, + 0.03213926777243614, + 0.021089905872941017, + 0.03325387462973595, + 0.013841336593031883, + 0.030931081622838974, + 0.007235029712319374, + 0.012991568073630333, + 0.007604224141687155, + 0.013897971250116825, + 0.005567467771470547, + 0.011600214056670666, + 0.004489446524530649, + 0.013379220850765705, + 0.013282429426908493, + 0.008000161498785019, + 0.02096511609852314, + 0.009075273759663105, + 0.012913879938423634, + 0.011643273755908012, + 0.013973657973110676, + 0.012675919570028782, + 0.006999157834798098, + 0.005373205989599228, + 0.012032384052872658, + 0.005269976332783699, + 0.006614150013774633, + 0.010834920220077038, + 0.009138135239481926, + 0.010868625715374947, + 0.0019443181809037924, + 0.004292106255888939, + 0.0033608493395149708, + 0.004953484050929546, + 0.002566966461017728, + 0.002217048080638051, + 0.0025808552745729685, + 0.002323311287909746, + 0.004116113763302565, + 0.00468696653842926, + 0.0041066501289606094, + 0.0042589385993778706, + 0.0028752456419169903, + 0.005851649213582277, + 0.002171550877392292, + 0.005740335676819086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04312288761138916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008250646642409265 + ], + "opposite_logprob_tokens": [ + -7.125824928283691 + ], + "logprob_reward": -0.0008250646642409265, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008250646642409265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19121573865413666, + 0.3210071921348572, + 0.0744144544005394, + 0.07181841880083084, + 0.08195088803768158, + 0.049253933131694794, + 0.08252127468585968, + 0.035097669810056686, + 0.028569214046001434, + 0.06415124237537384 + ], + "credit_map": [ + 0.01354973204433918, + 0.04742007330060005, + 0.018074996769428253, + 0.036651406437158585, + 0.020060788840055466, + 0.017955966293811798, + 0.01678384467959404, + 0.02071893773972988, + 0.03274336829781532, + 0.0659816563129425, + 0.04254242405295372, + 0.04583011567592621, + 0.03755471110343933, + 0.03798346593976021, + 0.025240184739232063, + 0.033131252974271774, + 0.00634756637737155, + 0.018038786947727203, + 0.008238679729402065, + 0.015231152065098286, + 0.0066467286087572575, + 0.006419455632567406, + 0.006088532507419586, + 0.007403550669550896, + 0.0058350027538836, + 0.01869240030646324, + 0.005658578593283892, + 0.013988330028951168, + 0.008373450487852097, + 0.0053771804086863995, + 0.0064508430659770966, + 0.0074426340870559216, + 0.007658451795578003, + 0.017108216881752014, + 0.007168253418058157, + 0.013556809164583683, + 0.01321824174374342, + 0.008538646623492241, + 0.006051880307495594, + 0.008650380186736584, + 0.007393865380436182, + 0.008788136765360832, + 0.004389737732708454, + 0.008743075653910637, + 0.005421028472483158, + 0.005196820478886366, + 0.0036028525792062283, + 0.005718417465686798, + 0.01008228212594986, + 0.010712174698710442, + 0.012813163921236992, + 0.010895850136876106, + 0.008529619313776493, + 0.00964315328747034, + 0.008207550272345543, + 0.011637476272881031, + 0.0033991015516221523, + 0.006653127260506153, + 0.004263649694621563, + 0.006485323887318373, + 0.004849849734455347, + 0.0032299247104674578, + 0.002918776124715805, + 0.0032979154493659735, + 0.0026554514188319445, + 0.004020280670374632, + 0.004229593090713024, + 0.004291000310331583, + 0.0037802744191139936, + 0.003163414541631937, + 0.0036381492391228676, + 0.0027910503558814526, + 0.00823313370347023, + 0.009547069668769836, + 0.00847022607922554, + 0.00864099059253931, + 0.005498083308339119, + 0.009559514932334423, + 0.006097119767218828, + 0.008105103857815266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0008250646642409265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032692551612854 + ], + "opposite_logprob_tokens": [ + -2.3503270149230957 + ], + "logprob_reward": -0.10032692551612854, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032692551612854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16846124827861786, + 0.1717611700296402, + 0.14942410588264465, + 0.06518697738647461, + 0.05182165279984474, + 0.14690351486206055, + 0.08940675109624863, + 0.06467951834201813, + 0.0457865446805954, + 0.046568553894758224 + ], + "credit_map": [ + 0.011725032702088356, + 0.029729334637522697, + 0.0189458429813385, + 0.033870916813611984, + 0.021111182868480682, + 0.018547696992754936, + 0.01334372628480196, + 0.021187523379921913, + 0.017042290419340134, + 0.03079335205256939, + 0.026874130591750145, + 0.019981004297733307, + 0.02291715517640114, + 0.01974036917090416, + 0.017425380647182465, + 0.016987493261694908, + 0.015945857390761375, + 0.032112967222929, + 0.010651734657585621, + 0.023899877443909645, + 0.014868389815092087, + 0.020558707416057587, + 0.012216176837682724, + 0.019170401617884636, + 0.005900925491005182, + 0.015509214252233505, + 0.007692266721278429, + 0.009388985112309456, + 0.0070402780547738075, + 0.006558873690664768, + 0.0058409119956195354, + 0.007255524396896362, + 0.005126888398081064, + 0.008630039170384407, + 0.0049876561388373375, + 0.010851943865418434, + 0.006432436406612396, + 0.005262027494609356, + 0.00398465059697628, + 0.006546011194586754, + 0.013911440037190914, + 0.026551391929388046, + 0.011393910273909569, + 0.02758847177028656, + 0.009889601729810238, + 0.021536702290177345, + 0.008784390985965729, + 0.027247607707977295, + 0.012084701098501682, + 0.008527967147529125, + 0.018465569242835045, + 0.008206810802221298, + 0.010431530885398388, + 0.008816794492304325, + 0.012631013058125973, + 0.010242369025945663, + 0.008231375366449356, + 0.011896633543074131, + 0.005934550892561674, + 0.011719975620508194, + 0.006554470397531986, + 0.007422344759106636, + 0.00477465707808733, + 0.008145510219037533, + 0.004485834389925003, + 0.006048132199794054, + 0.007160731125622988, + 0.004870903212577105, + 0.0048984321765601635, + 0.0067087942734360695, + 0.006439903751015663, + 0.005173815879970789, + 0.003093030536547303, + 0.01001300010830164, + 0.003412204561755061, + 0.012605048716068268, + 0.005075101740658283, + 0.0032451075967401266, + 0.004269985482096672, + 0.004855078179389238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.10032692551612854 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818707026541233 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818707026541233, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818707026541233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1797538697719574, + 0.1150282472372055, + 0.07351765781641006, + 0.10789918154478073, + 0.08847824484109879, + 0.15333077311515808, + 0.10628362745046616, + 0.04876834526658058, + 0.06405019760131836, + 0.06288983672857285 + ], + "credit_map": [ + 0.014123824425041676, + 0.04124579206109047, + 0.0242376159876585, + 0.024581801146268845, + 0.022315336391329765, + 0.016998374834656715, + 0.01772078312933445, + 0.018530340865254402, + 0.015751078724861145, + 0.023397296667099, + 0.010721719823777676, + 0.014150341041386127, + 0.00900685042142868, + 0.019058385863900185, + 0.007448219228535891, + 0.015494353137910366, + 0.006335489917546511, + 0.011987732723355293, + 0.008187863044440746, + 0.009542586281895638, + 0.01187349483370781, + 0.00879333820194006, + 0.00911227148026228, + 0.007684879470616579, + 0.009286535903811455, + 0.013243032619357109, + 0.016411783173680305, + 0.012580864131450653, + 0.014602551236748695, + 0.015960250049829483, + 0.014809475280344486, + 0.011004691012203693, + 0.008047458715736866, + 0.011929054744541645, + 0.011572073213756084, + 0.013833125121891499, + 0.010001958347856998, + 0.01404849998652935, + 0.00974580179899931, + 0.009300274774432182, + 0.014245269820094109, + 0.026218926534056664, + 0.010431678034365177, + 0.027256924659013748, + 0.010169170796871185, + 0.026709049940109253, + 0.007542448118329048, + 0.03075731359422207, + 0.015266742557287216, + 0.009142611175775528, + 0.021084118634462357, + 0.008392074145376682, + 0.013449741527438164, + 0.012358500622212887, + 0.013555163517594337, + 0.013034673407673836, + 0.004932055715471506, + 0.008087056688964367, + 0.006746831350028515, + 0.007683671545237303, + 0.005705312825739384, + 0.004480104427784681, + 0.006145495921373367, + 0.004987817723304033, + 0.006739792414009571, + 0.0077569917775690556, + 0.012773633003234863, + 0.007490614429116249, + 0.004879747051745653, + 0.008347246795892715, + 0.008150405250489712, + 0.007911765947937965, + 0.00427891593426466, + 0.006942782085388899, + 0.005795393604785204, + 0.008497311733663082, + 0.00794643722474575, + 0.010824877768754959, + 0.007106046658009291, + 0.011498071253299713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01818707026541233 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly advancing close-up is fixed on the screen of an old cathode-ray tube (CRT) monitor. The monitor's thick, beige plastic case, slightly yellowed by age, is covered in tiny scratches and dust. In the center of its slightly convex, bottomless black glass screen, a stick figure, composed entirely of bright, slightly glowing phosphorescent green vector lines, steps backward in a mechanical, clumsy manner. Its movements are not fluid animations, but rather a slow, staccato flow: one leg stiffly raises and lowers backward, and the other follows in an equally uncoordinated manner, its entire body flickering faintly and jumping between frames. This green figure exists isolated in the pure blackness of digital space, its cold glow the sole primary source of light, casting a retro and slightly mysterious aura upon its gloomy surroundings, creating a strange sense of contemplation upon the lonely operation of outdated technology.", + "seed": 7992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_12/index_5224_seed_7992_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023372726695379244, + "vlm_logprob": -0.023372726695379244, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.023372726695379244, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008095074445009232 + ], + "opposite_logprob_tokens": [ + -5.2580952644348145 + ], + "logprob_reward": -0.008095074445009232, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008095074445009232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08210170269012451, + 0.1181841641664505, + 0.07409105449914932, + 0.12570327520370483, + 0.039789214730262756, + 0.2981666028499603, + 0.05645042285323143, + 0.07680927217006683, + 0.05747793987393379, + 0.07122638076543808 + ], + "credit_map": [ + 0.009487154893577099, + 0.009152360260486603, + 0.007072231266647577, + 0.014945203438401222, + 0.01268845982849598, + 0.007660414557904005, + 0.01371200755238533, + 0.007383864838629961, + 0.017734400928020477, + 0.011637571267783642, + 0.02534172497689724, + 0.01409139297902584, + 0.011562428437173367, + 0.010233046486973763, + 0.014933498576283455, + 0.012650106102228165, + 0.010514521971344948, + 0.009875931777060032, + 0.005769048351794481, + 0.012813570909202099, + 0.0077656167559325695, + 0.006809730548411608, + 0.014645271003246307, + 0.0058973608538508415, + 0.015947913751006126, + 0.010730549693107605, + 0.030598290264606476, + 0.015980999916791916, + 0.010029873810708523, + 0.010840428993105888, + 0.01471086498349905, + 0.0168643519282341, + 0.006538541056215763, + 0.005306432954967022, + 0.0046919770538806915, + 0.0061636557802557945, + 0.0037438678555190563, + 0.003374634310603142, + 0.0066822599619627, + 0.0032878415659070015, + 0.04565350338816643, + 0.02440670318901539, + 0.08035494387149811, + 0.03832142427563667, + 0.020424125716090202, + 0.018572498112916946, + 0.03812219947576523, + 0.03231118619441986, + 0.008460955694317818, + 0.008296768181025982, + 0.006806232035160065, + 0.008640913292765617, + 0.0060543278232216835, + 0.005287546664476395, + 0.008589185774326324, + 0.004314495716243982, + 0.010740713216364384, + 0.008873149752616882, + 0.009522977285087109, + 0.01028834842145443, + 0.00704265758395195, + 0.008968062698841095, + 0.011880778707563877, + 0.009492580778896809, + 0.010761218145489693, + 0.0071846600621938705, + 0.006448464468121529, + 0.0076480950228869915, + 0.004473989363759756, + 0.006423517130315304, + 0.008746009320020676, + 0.005791989155113697, + 0.01000656932592392, + 0.006672316696494818, + 0.01334263477474451, + 0.008310225792229176, + 0.007418923079967499, + 0.007912172004580498, + 0.00782271008938551, + 0.009740821085870266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.008095074445009232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062120433896780014 + ], + "opposite_logprob_tokens": [ + -2.8121204376220703 + ], + "logprob_reward": -0.062120433896780014, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062120433896780014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05961272493004799, + 0.10164085775613785, + 0.02824307605624199, + 0.06527666747570038, + 0.037432711571455, + 0.4074419438838959, + 0.02000102773308754, + 0.07013837993144989, + 0.09525227546691895, + 0.11496040970087051 + ], + "credit_map": [ + 0.007909250445663929, + 0.006343176122754812, + 0.007078478578478098, + 0.011955957859754562, + 0.006166675593703985, + 0.005362468305975199, + 0.0107472138479352, + 0.004049491137266159, + 0.015437895432114601, + 0.011169823817908764, + 0.01965956762433052, + 0.012004679068922997, + 0.00848549697548151, + 0.010214803740382195, + 0.013413912616670132, + 0.011254667304456234, + 0.003479639533907175, + 0.0028227197472006083, + 0.0029196913819760084, + 0.0050201499834656715, + 0.002989011351019144, + 0.0017901244573295116, + 0.006868305150419474, + 0.0023534311912953854, + 0.009423846378922462, + 0.005301907658576965, + 0.016471674665808678, + 0.008406763896346092, + 0.005059199407696724, + 0.004933298099786043, + 0.00862623006105423, + 0.0070537361316382885, + 0.00443611852824688, + 0.004255776293575764, + 0.004356248304247856, + 0.007031700108200312, + 0.004084461368620396, + 0.0030162979383021593, + 0.0068919844925403595, + 0.0033601182512938976, + 0.06372028589248657, + 0.03420187160372734, + 0.10984336584806442, + 0.05137680098414421, + 0.026509450748562813, + 0.02602302096784115, + 0.050934288650751114, + 0.04483279585838318, + 0.0032174710649996996, + 0.0021321538370102644, + 0.0026247098576277494, + 0.003028529230505228, + 0.0020599309355020523, + 0.002005065092816949, + 0.0030841962434351444, + 0.0018489675130695105, + 0.00985738355666399, + 0.008463392965495586, + 0.010028962045907974, + 0.009163139387965202, + 0.006044619716703892, + 0.0076871998608112335, + 0.010318676009774208, + 0.008575000800192356, + 0.014662962406873703, + 0.014316265471279621, + 0.009011236019432545, + 0.014866771176457405, + 0.009783795103430748, + 0.010036708787083626, + 0.014154263772070408, + 0.008420255035161972, + 0.01350868958979845, + 0.011567567475140095, + 0.01744718663394451, + 0.017147570848464966, + 0.014622596092522144, + 0.0092522744089365, + 0.02008371241390705, + 0.011330800130963326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.062120433896780014 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0003979606262873858 + ], + "opposite_logprob_tokens": [ + -7.8753981590271 + ], + "logprob_reward": -0.0003979606262873858, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0003979606262873858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14426468312740326, + 0.1058480516076088, + 0.16588422656059265, + 0.12584491074085236, + 0.07686891406774521, + 0.06389550864696503, + 0.06651461869478226, + 0.050248801708221436, + 0.14673486351966858, + 0.053895413875579834 + ], + "credit_map": [ + 0.016441069543361664, + 0.013701308518648148, + 0.017723988741636276, + 0.02354288287460804, + 0.022069459781050682, + 0.010849527083337307, + 0.030350640416145325, + 0.009585808962583542, + 0.015736572444438934, + 0.011802030727267265, + 0.025449104607105255, + 0.012085525318980217, + 0.00940014235675335, + 0.008225420489907265, + 0.012541654519736767, + 0.010607593692839146, + 0.01770148240029812, + 0.018736721947789192, + 0.019872944802045822, + 0.030782731249928474, + 0.022451268509030342, + 0.011751360259950161, + 0.03216312453150749, + 0.012424597516655922, + 0.015158146619796753, + 0.011617208831012249, + 0.02703559584915638, + 0.01621428318321705, + 0.01193790603429079, + 0.011965102516114712, + 0.01654377393424511, + 0.015372895635664463, + 0.011198335327208042, + 0.009121478535234928, + 0.013501346111297607, + 0.009780186228454113, + 0.006330930162221193, + 0.007619740441441536, + 0.011506079696118832, + 0.007810812443494797, + 0.007749804761260748, + 0.006068961229175329, + 0.01569219119846821, + 0.010933063924312592, + 0.005235263612121344, + 0.002443687291815877, + 0.011371344327926636, + 0.004401196725666523, + 0.006530109792947769, + 0.006970501970499754, + 0.008939039893448353, + 0.011991003528237343, + 0.00691875908523798, + 0.005774880293756723, + 0.01306109968572855, + 0.006329221185296774, + 0.006817316170781851, + 0.004889883100986481, + 0.006647469010204077, + 0.008408923633396626, + 0.006048915442079306, + 0.0046107652597129345, + 0.008570648729801178, + 0.0042548831552267075, + 0.02185564860701561, + 0.01891169883310795, + 0.014345405623316765, + 0.020588679239153862, + 0.013985000550746918, + 0.01677354797720909, + 0.022599276155233383, + 0.017675606533885002, + 0.006146073807030916, + 0.005540736950933933, + 0.0067666564136743546, + 0.008100885897874832, + 0.007019418757408857, + 0.005148585885763168, + 0.009144613519310951, + 0.006028441712260246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.0003979606262873858 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370939567685127 + ], + "opposite_logprob_tokens": [ + -3.4087092876434326 + ], + "logprob_reward": -0.03370939567685127, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370939567685127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12437573075294495, + 0.0695587545633316, + 0.21852968633174896, + 0.06602662801742554, + 0.11486843228340149, + 0.16364312171936035, + 0.09151678532361984, + 0.046469688415527344, + 0.06991514563560486, + 0.03509604558348656 + ], + "credit_map": [ + 0.01679861731827259, + 0.0118577741086483, + 0.010064777918159962, + 0.021359050646424294, + 0.016825780272483826, + 0.011798173189163208, + 0.027295513078570366, + 0.008376042358577251, + 0.005765540990978479, + 0.00928807258605957, + 0.005880953744053841, + 0.011614760383963585, + 0.01179859321564436, + 0.005504566710442305, + 0.014539924450218678, + 0.005166338756680489, + 0.019856398925185204, + 0.023384608328342438, + 0.021200941875576973, + 0.038631077855825424, + 0.023851610720157623, + 0.016385357826948166, + 0.05796300619840622, + 0.01725670136511326, + 0.010802753269672394, + 0.00810222513973713, + 0.012014091946184635, + 0.00785155687481165, + 0.0054029179736971855, + 0.0063021001406013966, + 0.008052337914705276, + 0.007498648948967457, + 0.015230169519782066, + 0.01256287656724453, + 0.01637815497815609, + 0.015352222137153149, + 0.013828111812472343, + 0.011408703401684761, + 0.01917225494980812, + 0.010935939848423004, + 0.023262325674295425, + 0.013560216873884201, + 0.04563767462968826, + 0.024247456341981888, + 0.012549960054457188, + 0.007900802418589592, + 0.02252648025751114, + 0.013958199881017208, + 0.009275288321077824, + 0.009180848486721516, + 0.013164577074348927, + 0.017742251977324486, + 0.009045420214533806, + 0.006502601783722639, + 0.02030404657125473, + 0.006301748566329479, + 0.0035431301221251488, + 0.005121421534568071, + 0.004688202403485775, + 0.009269042871892452, + 0.007684892509132624, + 0.002999931341037154, + 0.01019979827105999, + 0.002963272389024496, + 0.010414504446089268, + 0.00826282799243927, + 0.0064292605966329575, + 0.009764189831912518, + 0.007124693598598242, + 0.007879149168729782, + 0.01188085600733757, + 0.008159664459526539, + 0.004660785663872957, + 0.0036814806517213583, + 0.0030002810526639223, + 0.005617326591163874, + 0.004054097458720207, + 0.003597319358959794, + 0.007072433829307556, + 0.003412319580093026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.03370939567685127 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of an old cathode-ray tube (CRT) monitor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the monitor's case thick, beige, and covered in tiny scratches and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stick figure composed of bright, glowing phosphorescent green vector lines visible on the screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the stick figure move backward in a mechanical, clumsy, staccato manner with stiff leg movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera fixed in a slow, steady close-up without panning or zooming?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540768831968307 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540768831968307, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540768831968307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17737475037574768, + 0.08197221904993057, + 0.07543092966079712, + 0.10333000868558884, + 0.09210500121116638, + 0.08423048257827759, + 0.1505914032459259, + 0.06733470410108566, + 0.11359387636184692, + 0.05403663218021393 + ], + "credit_map": [ + 0.022415153682231903, + 0.023151084780693054, + 0.013886976055800915, + 0.028844453394412994, + 0.02687397599220276, + 0.01944989152252674, + 0.02818630263209343, + 0.014566902071237564, + 0.010919783264398575, + 0.010981335304677486, + 0.006981649901717901, + 0.01279122568666935, + 0.010375681333243847, + 0.008230944164097309, + 0.015269069001078606, + 0.006422532722353935, + 0.009702512063086033, + 0.008340044878423214, + 0.010917683131992817, + 0.010982464998960495, + 0.007184009067714214, + 0.0069808876141905785, + 0.014519737102091312, + 0.006803593598306179, + 0.01748564839363098, + 0.010759184136986732, + 0.02150264009833336, + 0.01021895743906498, + 0.0069414954632520676, + 0.01015810202807188, + 0.011075998656451702, + 0.015187984332442284, + 0.013933568261563778, + 0.010236455127596855, + 0.0071740527637302876, + 0.015057288110256195, + 0.011978059075772762, + 0.008525291457772255, + 0.017310412600636482, + 0.007889868691563606, + 0.010250715538859367, + 0.010198083706200123, + 0.010098043829202652, + 0.013082113116979599, + 0.009808833710849285, + 0.00877873133867979, + 0.012407724745571613, + 0.00960623100399971, + 0.027389738708734512, + 0.021124064922332764, + 0.011650792323052883, + 0.020820898935198784, + 0.012849067337810993, + 0.017785873264074326, + 0.021208008751273155, + 0.017762957140803337, + 0.009473848156630993, + 0.007343770004808903, + 0.009081343188881874, + 0.009170085191726685, + 0.008743535727262497, + 0.006359767634421587, + 0.008983704261481762, + 0.008178647607564926, + 0.01702365092933178, + 0.014507696963846684, + 0.011611287482082844, + 0.016073042526841164, + 0.010264988988637924, + 0.01272234134376049, + 0.017594488337635994, + 0.013796386308968067, + 0.006656757555902004, + 0.0061929672956466675, + 0.006072297692298889, + 0.007558284793049097, + 0.005913992878049612, + 0.005226551555097103, + 0.01011844351887703, + 0.006297336891293526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.012540768831968307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_13/index_tensor([415, 415])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_13/index_tensor([415, 415])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..36b6d2adb13620ea42e29289a8c446509a0a65ed --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_13/index_tensor([415, 415])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping drone shot, from a bird's-eye view, slowly follows an ancient wall of weathered gray stone. It winds its way along a rugged ridge, stretching endlessly into the distance like the back of a giant dragon, until it disappears into the mist at the end of the sight. In the frame, a tiny figure, wearing a crimson windbreaker and carrying a bulging black backpack, advances steadily and rhythmically along the narrow, uneven stone path atop the wall. In the twilight hours, slanting golden sunlight bathes the scene in a warm, soft glow, casting the deep green mountains and the figures' long shadows across the ancient wall. The entire scene, set against the backdrop of this epic landscape, evokes a tranquil, contemplative atmosphere, imbued with the loneliness of personal exploration and the weight of history.", + "seed": 16167, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_13/index_415_seed_16167_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016631847340613605, + "vlm_logprob": -0.016631847340613605, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.016631847340613605, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030705302953720093 + ], + "opposite_logprob_tokens": [ + -4.030705451965332 + ], + "logprob_reward": -0.030705302953720093, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.030705302953720093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19551733136177063, + 0.06958112865686417, + 0.1888098567724228, + 0.10879166424274445, + 0.01680641807615757, + 0.049898501485586166, + 0.043424565345048904, + 0.19060927629470825, + 0.10046742856502533, + 0.03609386831521988 + ], + "credit_map": [ + 0.0190382432192564, + 0.020511679351329803, + 0.027466021478176117, + 0.03170686587691307, + 0.019332773983478546, + 0.019138580188155174, + 0.03168599680066109, + 0.026637159287929535, + 0.004443639889359474, + 0.007444278337061405, + 0.010182620026171207, + 0.008252299390733242, + 0.00893445871770382, + 0.013607287779450417, + 0.008607855997979641, + 0.008108689449727535, + 0.015180718153715134, + 0.041103627532720566, + 0.026018202304840088, + 0.03213629126548767, + 0.017475856468081474, + 0.02331666648387909, + 0.014522588811814785, + 0.01905590109527111, + 0.00788229238241911, + 0.010394309647381306, + 0.019819380715489388, + 0.009322847239673138, + 0.012938213534653187, + 0.03340338543057442, + 0.007074778433889151, + 0.007956461049616337, + 0.0017033759504556656, + 0.0018610249971970916, + 0.002715997863560915, + 0.0014743779320269823, + 0.0025721730198711157, + 0.0034924049396067858, + 0.0015260364161804318, + 0.0014610259095206857, + 0.007319734897464514, + 0.0050960141234099865, + 0.006844975985586643, + 0.00546454219147563, + 0.005578176584094763, + 0.00715365307405591, + 0.005568418651819229, + 0.00687298271805048, + 0.00429109251126647, + 0.005104731768369675, + 0.006413421593606472, + 0.00408570421859622, + 0.0047613936476409435, + 0.009310492314398289, + 0.005244592670351267, + 0.0042131394147872925, + 0.027102189138531685, + 0.018248174339532852, + 0.024473419412970543, + 0.022495299577713013, + 0.027051405981183052, + 0.034497521817684174, + 0.02355930767953396, + 0.013181962072849274, + 0.015961438417434692, + 0.008471090346574783, + 0.0340907908976078, + 0.012237523682415485, + 0.00736164441332221, + 0.010156113654375076, + 0.006852089427411556, + 0.00533673819154501, + 0.003908223006874323, + 0.005400072783231735, + 0.008234364911913872, + 0.0041004871018230915, + 0.0040178787894546986, + 0.006201968528330326, + 0.002250547055155039, + 0.001980326371267438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.030705302953720093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018208960071206093 + ], + "opposite_logprob_tokens": [ + -4.018208980560303 + ], + "logprob_reward": -0.018208960071206093, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018208960071206093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20770850777626038, + 0.03803228214383125, + 0.03904297202825546, + 0.060136497020721436, + 0.029692288488149643, + 0.09578802436590195, + 0.10524091124534607, + 0.052876055240631104, + 0.31748273968696594, + 0.05399973317980766 + ], + "credit_map": [ + 0.015498751774430275, + 0.017104286700487137, + 0.028612947091460228, + 0.028380012139678, + 0.024211373180150986, + 0.020043684169650078, + 0.038829125463962555, + 0.03502831980586052, + 0.0025050428230315447, + 0.005893462337553501, + 0.005426897667348385, + 0.004830167628824711, + 0.0033971017692238092, + 0.006576969753950834, + 0.0034628240391612053, + 0.0059398142620921135, + 0.003810599911957979, + 0.0073289815336465836, + 0.005060469266027212, + 0.005685463547706604, + 0.0036343049723654985, + 0.005074806045740843, + 0.0031503261998295784, + 0.005298015661537647, + 0.00410214951261878, + 0.007076109293848276, + 0.01073384191840887, + 0.004974238574504852, + 0.006352443713694811, + 0.01732342131435871, + 0.004695770796388388, + 0.004878520499914885, + 0.003028930863365531, + 0.003031299449503422, + 0.005446924827992916, + 0.0027837117668241262, + 0.0033125218469649553, + 0.00645453529432416, + 0.0030697535257786512, + 0.002564607886597514, + 0.010253057815134525, + 0.010330111719667912, + 0.012084482237696648, + 0.01513851061463356, + 0.008443297818303108, + 0.013954191468656063, + 0.014353785663843155, + 0.011230584233999252, + 0.01034354418516159, + 0.008078193292021751, + 0.016959937289357185, + 0.009591402485966682, + 0.013390690088272095, + 0.023429024964571, + 0.012345163151621819, + 0.011102951131761074, + 0.010528990998864174, + 0.004390611778944731, + 0.014477909542620182, + 0.004300265107303858, + 0.0031255849171429873, + 0.008596442639827728, + 0.004419393837451935, + 0.0030368552543222904, + 0.04828618839383125, + 0.0248295608907938, + 0.06850514560937881, + 0.042997803539037704, + 0.03257092088460922, + 0.043324802070856094, + 0.03567056730389595, + 0.021297749131917953, + 0.010896083898842335, + 0.003469330258667469, + 0.011884519830346107, + 0.004077771212905645, + 0.005469590425491333, + 0.01069044042378664, + 0.0036243628710508347, + 0.0038876349572092295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.018208960071206093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011797142215073109 + ], + "opposite_logprob_tokens": [ + -6.7511796951293945 + ], + "logprob_reward": -0.0011797142215073109, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011797142215073109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2758011221885681, + 0.12129402905702591, + 0.09055077284574509, + 0.049380652606487274, + 0.03426510468125343, + 0.087483711540699, + 0.04773849621415138, + 0.1203140914440155, + 0.12439816445112228, + 0.04877384379506111 + ], + "credit_map": [ + 0.03728843107819557, + 0.04576198756694794, + 0.04004789888858795, + 0.04576408863067627, + 0.027112970128655434, + 0.021342983469367027, + 0.03312643989920616, + 0.025356296449899673, + 0.009096688590943813, + 0.023129666224122047, + 0.019818810746073723, + 0.025285761803388596, + 0.009829377755522728, + 0.015955131500959396, + 0.007656568195670843, + 0.0105220265686512, + 0.005656261462718248, + 0.021071163937449455, + 0.01385174784809351, + 0.015187624841928482, + 0.007582733873277903, + 0.013006889261305332, + 0.0057640401646494865, + 0.008430315181612968, + 0.005911676678806543, + 0.006752260960638523, + 0.008929293602705002, + 0.004995156545192003, + 0.00382433389313519, + 0.010958020575344563, + 0.004446346778422594, + 0.0035635647363960743, + 0.0051633394323289394, + 0.0044054901227355, + 0.005008965265005827, + 0.003735974431037903, + 0.0029069397132843733, + 0.005826150998473167, + 0.003476498881354928, + 0.003741747234016657, + 0.013482610695064068, + 0.007692570798099041, + 0.021077608689665794, + 0.011816557496786118, + 0.006582781206816435, + 0.011384902521967888, + 0.007861507125198841, + 0.0075851683504879475, + 0.007220293860882521, + 0.004083716310560703, + 0.012863210402429104, + 0.003628472099080682, + 0.003237457713112235, + 0.008412940427660942, + 0.00451832776889205, + 0.003774080891162157, + 0.015658708289265633, + 0.012398459948599339, + 0.015432542189955711, + 0.015019308775663376, + 0.017271876335144043, + 0.020520761609077454, + 0.01541764847934246, + 0.008594781160354614, + 0.025375785306096077, + 0.007834396325051785, + 0.047064900398254395, + 0.010489852167665958, + 0.007905863225460052, + 0.011220637708902359, + 0.0069620683789253235, + 0.007544663269072771, + 0.005734538193792105, + 0.007843157276511192, + 0.009658602997660637, + 0.007465225178748369, + 0.003935378976166248, + 0.006176608614623547, + 0.004359567537903786, + 0.003600766882300377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0011797142215073109 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328609488904476 + ], + "opposite_logprob_tokens": [ + -3.7732861042022705 + ], + "logprob_reward": -0.02328609488904476, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328609488904476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33931979537010193, + 0.060753095895051956, + 0.03180587291717529, + 0.03686722368001938, + 0.029545912519097328, + 0.24551455676555634, + 0.0841297060251236, + 0.1103752925992012, + 0.045867908746004105, + 0.015820631757378578 + ], + "credit_map": [ + 0.01945318654179573, + 0.026035649701952934, + 0.04918242245912552, + 0.04908646270632744, + 0.04291928932070732, + 0.032304547727108, + 0.06480716168880463, + 0.055531080812215805, + 0.0036445739679038525, + 0.005359597038477659, + 0.009046167135238647, + 0.00652401614934206, + 0.007753930985927582, + 0.012536388821899891, + 0.007221792358905077, + 0.0086666289716959, + 0.002087170025333762, + 0.0042233094573020935, + 0.004962797276675701, + 0.005056224763393402, + 0.003755484474822879, + 0.005318029318004847, + 0.0030678333714604378, + 0.0033350251615047455, + 0.003809839254245162, + 0.0044630141928792, + 0.005201900843530893, + 0.006405777297914028, + 0.0033799412194639444, + 0.007843141444027424, + 0.0031614615581929684, + 0.002602145541459322, + 0.00351652386598289, + 0.003695820923894644, + 0.0045493412762880325, + 0.0036241980269551277, + 0.002770921913906932, + 0.0045124622993171215, + 0.0033001259434968233, + 0.0035765166394412518, + 0.03046238236129284, + 0.023103630170226097, + 0.05718982219696045, + 0.03464651480317116, + 0.017106076702475548, + 0.033799950033426285, + 0.026241013780236244, + 0.022965166717767715, + 0.014676648192107677, + 0.004951717332005501, + 0.02723071165382862, + 0.005541960243135691, + 0.004655796103179455, + 0.015843819826841354, + 0.005708824843168259, + 0.005520225502550602, + 0.021836716681718826, + 0.0076859258115291595, + 0.03882024437189102, + 0.008384312503039837, + 0.007403908297419548, + 0.0157832819968462, + 0.006038588937371969, + 0.004422304220497608, + 0.009542444720864296, + 0.00283321482129395, + 0.01691397652029991, + 0.0041915797628462315, + 0.0031063067726790905, + 0.004300398752093315, + 0.002614639000967145, + 0.002365349791944027, + 0.003403328824788332, + 0.0018680266803130507, + 0.0037888868246227503, + 0.0019337893463671207, + 0.0013247885508462787, + 0.0013837238075211644, + 0.0012950898380950093, + 0.0008229991071857512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.02328609488904476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977916456758976 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.00977916456758976, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977916456758976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20599940419197083, + 0.10796802490949631, + 0.1664009541273117, + 0.05246352031826973, + 0.04852445796132088, + 0.057476967573165894, + 0.13537392020225525, + 0.041626639664173126, + 0.13778258860111237, + 0.0463835746049881 + ], + "credit_map": [ + 0.014303252100944519, + 0.022081049159169197, + 0.029746437445282936, + 0.030792275443673134, + 0.023859063163399696, + 0.020785633474588394, + 0.03448692336678505, + 0.029944756999611855, + 0.008947842754423618, + 0.016157614067196846, + 0.015601771883666515, + 0.013553818687796593, + 0.010514458641409874, + 0.02210182324051857, + 0.008477278053760529, + 0.01261339895427227, + 0.01789286360144615, + 0.030056428164243698, + 0.021785205230116844, + 0.024488821625709534, + 0.015945112332701683, + 0.02402270771563053, + 0.012476908974349499, + 0.019732890650629997, + 0.004676776472479105, + 0.00571798300370574, + 0.008920061402022839, + 0.004406248219311237, + 0.0054075950756669044, + 0.01358809508383274, + 0.00435103103518486, + 0.005395727232098579, + 0.0035542165860533714, + 0.0063588484190404415, + 0.007006751839071512, + 0.005363950971513987, + 0.008970681577920914, + 0.008806923404335976, + 0.004027354530990124, + 0.0044357250444591045, + 0.0061359587125480175, + 0.006021152716130018, + 0.012058586813509464, + 0.005210359115153551, + 0.008191347122192383, + 0.008662943728268147, + 0.005506030283868313, + 0.00569058395922184, + 0.013217037543654442, + 0.012715969234704971, + 0.017799679189920425, + 0.014851742424070835, + 0.02073352225124836, + 0.017051922157406807, + 0.0246028583496809, + 0.014401169493794441, + 0.004942874889820814, + 0.00421965541318059, + 0.007493217010051012, + 0.005347512662410736, + 0.0048573859967291355, + 0.00812622718513012, + 0.0037940179463475943, + 0.0028457422740757465, + 0.02337595447897911, + 0.013000219129025936, + 0.046409331262111664, + 0.015074249356985092, + 0.009589606896042824, + 0.013081571087241173, + 0.009950262494385242, + 0.007301385048776865, + 0.005980186630040407, + 0.0039219241589307785, + 0.008447832427918911, + 0.005871035158634186, + 0.004939509555697441, + 0.007492192555218935, + 0.0058329300954937935, + 0.0038979623932391405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.00977916456758976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping drone shot, from a bird's-eye view, slowly follows an ancient wall of weathered gray stone. It winds its way along a rugged ridge, stretching endlessly into the distance like the back of a giant dragon, until it disappears into the mist at the end of the sight. In the frame, a tiny figure, wearing a crimson windbreaker and carrying a bulging black backpack, advances steadily and rhythmically along the narrow, uneven stone path atop the wall. In the twilight hours, slanting golden sunlight bathes the scene in a warm, soft glow, casting the deep green mountains and the figures' long shadows across the ancient wall. The entire scene, set against the backdrop of this epic landscape, evokes a tranquil, contemplative atmosphere, imbued with the loneliness of personal exploration and the weight of history.", + "seed": 16168, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_13/index_415_seed_16168_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06950355363078416, + "vlm_logprob": -0.06950355363078416, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.06950355363078416, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012696034274995327 + ], + "opposite_logprob_tokens": [ + -5.137696266174316 + ], + "logprob_reward": -0.012696034274995327, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.012696034274995327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17608456313610077, + 0.10208257287740707, + 0.029508668929338455, + 0.16430632770061493, + 0.026034899055957794, + 0.054354943335056305, + 0.28425219655036926, + 0.06793804466724396, + 0.07407896220684052, + 0.021358821541070938 + ], + "credit_map": [ + 0.020513899624347687, + 0.018276356160640717, + 0.021367762237787247, + 0.015922104939818382, + 0.01763378456234932, + 0.033647287636995316, + 0.022774245589971542, + 0.025949114933609962, + 0.021010471507906914, + 0.007985437288880348, + 0.016837509348988533, + 0.008618656545877457, + 0.010785441845655441, + 0.02034599334001541, + 0.008046812377870083, + 0.008452249690890312, + 0.005418058950453997, + 0.0024605838116258383, + 0.005668409168720245, + 0.003264976665377617, + 0.002221526810899377, + 0.004500750917941332, + 0.0026173018850386143, + 0.0033570611849427223, + 0.015735890716314316, + 0.009386766701936722, + 0.017551014199852943, + 0.00830149371176958, + 0.004670728463679552, + 0.051284100860357285, + 0.009803984314203262, + 0.047572363168001175, + 0.0047872173599898815, + 0.002338463207706809, + 0.004737663548439741, + 0.0019024586072191596, + 0.002657619072124362, + 0.004238496534526348, + 0.002196085173636675, + 0.0031768977642059326, + 0.006937412079423666, + 0.0045508588664233685, + 0.008297175168991089, + 0.0040736873634159565, + 0.003707915311679244, + 0.012187894433736801, + 0.004407972563058138, + 0.010192027315497398, + 0.025621704757213593, + 0.02305362932384014, + 0.09281522780656815, + 0.018700456246733665, + 0.015108590945601463, + 0.08765634149312973, + 0.010939897038042545, + 0.010356339626014233, + 0.0076646143570542336, + 0.007215334102511406, + 0.014505261555314064, + 0.005868827924132347, + 0.0050308299250900745, + 0.01136730145663023, + 0.008197803050279617, + 0.008088071830570698, + 0.006099933758378029, + 0.0022765130270272493, + 0.02339348942041397, + 0.004439919255673885, + 0.0035477622877806425, + 0.02749384567141533, + 0.003105456940829754, + 0.0037220388185232878, + 0.0026665993500500917, + 0.0018113268306478858, + 0.00265789614059031, + 0.002154488814994693, + 0.002471547806635499, + 0.004131297580897808, + 0.0021530420053750277, + 0.003312622895464301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.012696034274995327 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813558280467987 + ], + "opposite_logprob_tokens": [ + -1.406355857849121 + ], + "logprob_reward": -0.2813558280467987, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813558280467987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19381915032863617, + 0.11546029895544052, + 0.03265116736292839, + 0.057385314255952835, + 0.02932320348918438, + 0.07788510620594025, + 0.2927074432373047, + 0.13630062341690063, + 0.044408995658159256, + 0.020058704540133476 + ], + "credit_map": [ + 0.025564443320035934, + 0.023236243054270744, + 0.03032422438263893, + 0.018282312899827957, + 0.021944483742117882, + 0.03597118332982063, + 0.017242755740880966, + 0.021253498271107674, + 0.011689522303640842, + 0.01315789483487606, + 0.020498095080256462, + 0.012513023801147938, + 0.009379139170050621, + 0.01697588339447975, + 0.01768241450190544, + 0.013564327731728554, + 0.0049903132021427155, + 0.003137618536129594, + 0.007995223626494408, + 0.003316935384646058, + 0.002361137652769685, + 0.004671619739383459, + 0.0031546654645353556, + 0.003023654455319047, + 0.007217186037451029, + 0.004881557077169418, + 0.009387818165123463, + 0.004246214404702187, + 0.00417501712217927, + 0.0108792120590806, + 0.0062273419462144375, + 0.010370972566306591, + 0.004685398191213608, + 0.002993560628965497, + 0.00563669903203845, + 0.0024713200982660055, + 0.0023548565804958344, + 0.004670142661780119, + 0.0030243354849517345, + 0.003486890811473131, + 0.008042445406317711, + 0.0059111034497618675, + 0.013857943005859852, + 0.005809118505567312, + 0.004907159600406885, + 0.02062065899372101, + 0.0055707162246108055, + 0.013165962882339954, + 0.02776811271905899, + 0.02488754689693451, + 0.08993509411811829, + 0.021508267149329185, + 0.017156651243567467, + 0.08361586928367615, + 0.015122879296541214, + 0.012713000178337097, + 0.020083241164684296, + 0.010995778255164623, + 0.02114207111299038, + 0.011745928786695004, + 0.01378458272665739, + 0.033187974244356155, + 0.013728920370340347, + 0.011632129549980164, + 0.0043411231599748135, + 0.0015997366281226277, + 0.0133826844394207, + 0.002715991111472249, + 0.0020363598596304655, + 0.016146406531333923, + 0.0019453945569694042, + 0.002241299021989107, + 0.0022667585872113705, + 0.0012584523065015674, + 0.003421455156058073, + 0.002258836291730404, + 0.001367272692732513, + 0.0027782812248915434, + 0.0033203475177288055, + 0.0033872986678034067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.2813558280467987 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006338732782751322 + ], + "opposite_logprob_tokens": [ + -7.375633716583252 + ], + "logprob_reward": -0.0006338732782751322, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006338732782751322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11973312497138977, + 0.10222160071134567, + 0.07430700212717056, + 0.1542554795742035, + 0.06225862354040146, + 0.0241007711738348, + 0.1907445341348648, + 0.11020847409963608, + 0.13113722205162048, + 0.03103323094546795 + ], + "credit_map": [ + 0.018937593325972557, + 0.01261996477842331, + 0.016684722155332565, + 0.012213246896862984, + 0.0158078633248806, + 0.017256848514080048, + 0.014837157912552357, + 0.011375722475349903, + 0.015866002067923546, + 0.009388965554535389, + 0.016135361045598984, + 0.008743824437260628, + 0.007182934787124395, + 0.019341666251420975, + 0.014064863324165344, + 0.011497965082526207, + 0.010130728594958782, + 0.0069778235629200935, + 0.016612621024250984, + 0.006913033314049244, + 0.005511221010237932, + 0.010490559041500092, + 0.009332841262221336, + 0.008338174782693386, + 0.01339145191013813, + 0.009208607487380505, + 0.01755315624177456, + 0.00836318451911211, + 0.006425393745303154, + 0.043554261326789856, + 0.011604578234255314, + 0.04415483400225639, + 0.010896764695644379, + 0.005722604226320982, + 0.011551612056791782, + 0.004551497288048267, + 0.004059735685586929, + 0.010008208453655243, + 0.007936800830066204, + 0.007531393319368362, + 0.004173874389380217, + 0.002776330104097724, + 0.004148086067289114, + 0.002319751074537635, + 0.0016019220929592848, + 0.0035483618266880512, + 0.002567416289821267, + 0.0029650270007550716, + 0.016895612701773643, + 0.016333194449543953, + 0.0669812336564064, + 0.0106124859303236, + 0.010716116055846214, + 0.0556306429207325, + 0.007225602399557829, + 0.006349637173116207, + 0.010940372943878174, + 0.011869237758219242, + 0.035071682184934616, + 0.010565181262791157, + 0.005667050369083881, + 0.020871499553322792, + 0.008010799996554852, + 0.0072126430459320545, + 0.012071828357875347, + 0.0032667224295437336, + 0.04179376736283302, + 0.00837687961757183, + 0.0057581025175750256, + 0.05046870931982994, + 0.004727560561150312, + 0.0046736495569348335, + 0.004404333885759115, + 0.002057361416518688, + 0.0043856254778802395, + 0.003130245953798294, + 0.002305758884176612, + 0.006098490674048662, + 0.003212802344933152, + 0.005438611377030611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.0006338732782751322 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305369406938553 + ], + "opposite_logprob_tokens": [ + -3.16805362701416 + ], + "logprob_reward": -0.04305369406938553, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305369406938553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21508626639842987, + 0.13065990805625916, + 0.04497220367193222, + 0.11745565384626389, + 0.03952888026833534, + 0.01833736151456833, + 0.27525386214256287, + 0.06164698675274849, + 0.08118831366300583, + 0.01587062142789364 + ], + "credit_map": [ + 0.04353230446577072, + 0.022854033857584, + 0.02984614297747612, + 0.019286172464489937, + 0.020814532414078712, + 0.03336857631802559, + 0.022550128400325775, + 0.022834373638033867, + 0.02173101156949997, + 0.01268034242093563, + 0.019530633464455605, + 0.013662260957062244, + 0.011329042725265026, + 0.025141796097159386, + 0.012085441499948502, + 0.014499370940029621, + 0.00795783568173647, + 0.004030916839838028, + 0.008422798477113247, + 0.004301247652620077, + 0.00376787013374269, + 0.007477923762053251, + 0.004978319630026817, + 0.0040352921932935715, + 0.01010210532695055, + 0.007045346777886152, + 0.01372181810438633, + 0.005513403564691544, + 0.004443751182407141, + 0.03493952006101608, + 0.007674090098589659, + 0.03401562198996544, + 0.006777655798941851, + 0.00395126361399889, + 0.006725257262587547, + 0.0032160289119929075, + 0.004742254503071308, + 0.006204533390700817, + 0.003990178927779198, + 0.003921709954738617, + 0.002433107001706958, + 0.001951451413333416, + 0.002903594169765711, + 0.001843897858634591, + 0.0013820503372699022, + 0.002526740776374936, + 0.0030046780593693256, + 0.002291840733960271, + 0.023559223860502243, + 0.023000303655862808, + 0.0972311794757843, + 0.01571224257349968, + 0.015246354043483734, + 0.08209216594696045, + 0.009618986397981644, + 0.008793400600552559, + 0.007033882196992636, + 0.004926276858896017, + 0.015249256044626236, + 0.006401482503861189, + 0.004949825815856457, + 0.014837168157100677, + 0.00419890321791172, + 0.004050193354487419, + 0.007491621654480696, + 0.0017747372621670365, + 0.024755433201789856, + 0.004507300443947315, + 0.0035008536651730537, + 0.032362449914216995, + 0.003650474362075329, + 0.0031454397831112146, + 0.002331796567887068, + 0.001931533683091402, + 0.0018382909474894404, + 0.001587010221555829, + 0.0014475809875875711, + 0.0023849140852689743, + 0.00191529153380543, + 0.0024342029355466366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.04305369406938553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778338484466076 + ], + "opposite_logprob_tokens": [ + -4.6347784996032715 + ], + "logprob_reward": -0.009778338484466076, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778338484466076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12045756727457047, + 0.1269785463809967, + 0.037536732852458954, + 0.04920430853962898, + 0.04037824273109436, + 0.045891981571912766, + 0.3213675320148468, + 0.06456943601369858, + 0.1797105073928833, + 0.013905134983360767 + ], + "credit_map": [ + 0.01592205837368965, + 0.015495248138904572, + 0.014914841391146183, + 0.012640729546546936, + 0.0160063486546278, + 0.014342728070914745, + 0.01722743920981884, + 0.013908177614212036, + 0.01746853068470955, + 0.012447118759155273, + 0.017071140930056572, + 0.01567230373620987, + 0.009491642005741596, + 0.020624838769435883, + 0.01919492334127426, + 0.015008058398962021, + 0.004369694739580154, + 0.004725351929664612, + 0.005128690507262945, + 0.004553965758532286, + 0.0029770140536129475, + 0.005669526290148497, + 0.0058860513381659985, + 0.004226439166814089, + 0.006106243468821049, + 0.00488918274641037, + 0.007764683570712805, + 0.0034319977276027203, + 0.004158377647399902, + 0.0079200379550457, + 0.004755079280585051, + 0.010178711265325546, + 0.004011171404272318, + 0.004085914697498083, + 0.006426705047488213, + 0.0036077615804970264, + 0.004171170759946108, + 0.006078177597373724, + 0.0064383926801383495, + 0.005558950360864401, + 0.0057411896996200085, + 0.003526103449985385, + 0.006553521379828453, + 0.0028690528124570847, + 0.0028911754488945007, + 0.010722503997385502, + 0.003298637457191944, + 0.010289797559380531, + 0.028539177030324936, + 0.026559550315141678, + 0.10968165844678879, + 0.019202275201678276, + 0.017483705654740334, + 0.09789656102657318, + 0.01116210501641035, + 0.010842559859156609, + 0.006839924957603216, + 0.006567634642124176, + 0.01596989296376705, + 0.007510175462812185, + 0.005187587812542915, + 0.01023352425545454, + 0.0059820180758833885, + 0.006278681103140116, + 0.015464898198843002, + 0.004800061695277691, + 0.05646303668618202, + 0.011731992475688457, + 0.00796809047460556, + 0.06907933950424194, + 0.0077800326980650425, + 0.0064230673015117645, + 0.0018195874290540814, + 0.0013775753322988749, + 0.0016493098810315132, + 0.001513971365056932, + 0.0011900874087587, + 0.002278550760820508, + 0.0018392314668744802, + 0.002236823085695505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.009778338484466076 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping drone shot, from a bird's-eye view, slowly follows an ancient wall of weathered gray stone. It winds its way along a rugged ridge, stretching endlessly into the distance like the back of a giant dragon, until it disappears into the mist at the end of the sight. In the frame, a tiny figure, wearing a crimson windbreaker and carrying a bulging black backpack, advances steadily and rhythmically along the narrow, uneven stone path atop the wall. In the twilight hours, slanting golden sunlight bathes the scene in a warm, soft glow, casting the deep green mountains and the figures' long shadows across the ancient wall. The entire scene, set against the backdrop of this epic landscape, evokes a tranquil, contemplative atmosphere, imbued with the loneliness of personal exploration and the weight of history.", + "seed": 16169, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_13/index_415_seed_16169_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05643040401628241, + "vlm_logprob": -0.05643040401628241, + "vlm_yes_no_margin": 4.099999952316284, + "vlm_token_credit": -0.05643040401628241, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05058731511235237 + ], + "opposite_logprob_tokens": [ + -3.3005874156951904 + ], + "logprob_reward": -0.05058731511235237, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05058731511235237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3764071762561798, + 0.09719803929328918, + 0.06818914413452148, + 0.051309067755937576, + 0.12948447465896606, + 0.08578752726316452, + 0.0882183164358139, + 0.03338172659277916, + 0.02630118653178215, + 0.04372328892350197 + ], + "credit_map": [ + 0.025388969108462334, + 0.03352675214409828, + 0.06112350896000862, + 0.044543128460645676, + 0.03645998239517212, + 0.05737775191664696, + 0.03618963807821274, + 0.08179745078086853, + 0.008023678325116634, + 0.009786293841898441, + 0.020864704623818398, + 0.01001288928091526, + 0.00877709873020649, + 0.010930806398391724, + 0.014454134739935398, + 0.014348437078297138, + 0.005202845670282841, + 0.00782687310129404, + 0.013391154818236828, + 0.008491800166666508, + 0.0072912974283099174, + 0.008539278991520405, + 0.0074903275817632675, + 0.009955567307770252, + 0.004174772650003433, + 0.007593509741127491, + 0.007435915060341358, + 0.006936278194189072, + 0.005373056512326002, + 0.005659831687808037, + 0.006734264548867941, + 0.007401441689580679, + 0.007721113972365856, + 0.016652055084705353, + 0.01589088700711727, + 0.012182535603642464, + 0.014814674854278564, + 0.011886751279234886, + 0.03153626620769501, + 0.018800200894474983, + 0.005728614050894976, + 0.009123318828642368, + 0.01349656842648983, + 0.00757596967741847, + 0.009213621728122234, + 0.010535168461501598, + 0.017691917717456818, + 0.012422355823218822, + 0.007351779378950596, + 0.012751249596476555, + 0.01319847721606493, + 0.008982907049357891, + 0.008491483516991138, + 0.005821771454066038, + 0.018256064504384995, + 0.013364586047828197, + 0.002716892631724477, + 0.00273826252669096, + 0.004221929702907801, + 0.0031993892043828964, + 0.002759744878858328, + 0.0054413313046097755, + 0.006681927014142275, + 0.005622252821922302, + 0.0022698198445141315, + 0.0032720211893320084, + 0.0041906521655619144, + 0.002474240493029356, + 0.0033510897774249315, + 0.00356114050373435, + 0.0033662773203104734, + 0.0038159466348588467, + 0.004763856530189514, + 0.004166931379586458, + 0.006397696677595377, + 0.004695599898695946, + 0.004719859920442104, + 0.004583364352583885, + 0.007737141102552414, + 0.0066588385961949825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.05058731511235237 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018210599198937416 + ], + "opposite_logprob_tokens": [ + -4.018210411071777 + ], + "logprob_reward": -0.018210599198937416, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018210599198937416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06820815801620483, + 0.03666028380393982, + 0.03029652126133442, + 0.07511115074157715, + 0.1201552152633667, + 0.15201601386070251, + 0.056170858442783356, + 0.10215632617473602, + 0.046909451484680176, + 0.3123159408569336 + ], + "credit_map": [ + 0.005271546076983213, + 0.005501388572156429, + 0.011299476027488708, + 0.008154718205332756, + 0.007046821061521769, + 0.013930784538388252, + 0.00786446500569582, + 0.009138960391283035, + 0.0028124935925006866, + 0.004198556765913963, + 0.006449830252677202, + 0.005299619864672422, + 0.003895085770636797, + 0.0043664611876010895, + 0.004179516341537237, + 0.0054587204940617085, + 0.0019609518349170685, + 0.0037051322869956493, + 0.003570115426555276, + 0.0022867422085255384, + 0.003167665796354413, + 0.004160802345722914, + 0.00692814402282238, + 0.004516966640949249, + 0.005349916871637106, + 0.01246700156480074, + 0.00930772814899683, + 0.006132860202342272, + 0.006760230753570795, + 0.006467271130532026, + 0.017180515453219414, + 0.011445626616477966, + 0.007952937856316566, + 0.011071092449128628, + 0.014433538541197777, + 0.010747967287898064, + 0.011855999939143658, + 0.013236049562692642, + 0.0307651124894619, + 0.020092515274882317, + 0.009975151158869267, + 0.019273726269602776, + 0.018468305468559265, + 0.014836656861007214, + 0.01345811877399683, + 0.011293081566691399, + 0.03736882284283638, + 0.027342163026332855, + 0.00467293756082654, + 0.006020458415150642, + 0.010186088271439075, + 0.005708110518753529, + 0.0058214059099555016, + 0.004867499694228172, + 0.010362798348069191, + 0.008531558327376842, + 0.005817017983645201, + 0.007687011267989874, + 0.013829672709107399, + 0.010289530269801617, + 0.00941828079521656, + 0.01857038401067257, + 0.018499432131648064, + 0.018045000731945038, + 0.0036771055310964584, + 0.003630286082625389, + 0.007871969603002071, + 0.00462528457865119, + 0.006541421636939049, + 0.00709190871566534, + 0.006174428854137659, + 0.007297048345208168, + 0.01701958477497101, + 0.030504129827022552, + 0.0480596087872982, + 0.02785128355026245, + 0.030447259545326233, + 0.04344598948955536, + 0.06332331895828247, + 0.05166476219892502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.018210599198937416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007198841194622219 + ], + "opposite_logprob_tokens": [ + -7.250720024108887 + ], + "logprob_reward": -0.0007198841194622219, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007198841194622219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39460188150405884, + 0.13800667226314545, + 0.09306760132312775, + 0.048679836094379425, + 0.1045522391796112, + 0.059386517852544785, + 0.0582377053797245, + 0.036611393094062805, + 0.024957988411188126, + 0.04189814627170563 + ], + "credit_map": [ + 0.025518201291561127, + 0.031103316694498062, + 0.07144714891910553, + 0.03948231041431427, + 0.03608844429254532, + 0.05596557632088661, + 0.04215909168124199, + 0.09283777326345444, + 0.00796972494572401, + 0.012453816831111908, + 0.02908034808933735, + 0.011821300722658634, + 0.011665420606732368, + 0.019785575568675995, + 0.020187171176075935, + 0.025043310597538948, + 0.007828081026673317, + 0.010992375202476978, + 0.012584153562784195, + 0.009987832978367805, + 0.008235931396484375, + 0.011085046455264091, + 0.01826900988817215, + 0.014085161499679089, + 0.003964639268815517, + 0.006942638661712408, + 0.007058634888380766, + 0.005486896727234125, + 0.0053634983487427235, + 0.005761881358921528, + 0.0068685561418533325, + 0.0072330874390900135, + 0.006135701667517424, + 0.009191730059683323, + 0.015810975804924965, + 0.0083710802718997, + 0.013067960739135742, + 0.011546109803020954, + 0.025793636217713356, + 0.014635035768151283, + 0.003727055387571454, + 0.0054033054038882256, + 0.00927969254553318, + 0.0051193321123719215, + 0.006500015966594219, + 0.007050597108900547, + 0.01399664580821991, + 0.008309870958328247, + 0.006069735623896122, + 0.008421359583735466, + 0.009117343463003635, + 0.006691007874906063, + 0.006657974794507027, + 0.005708320066332817, + 0.006909529212862253, + 0.008662433363497257, + 0.0028504100628197193, + 0.0039212643168866634, + 0.004739841446280479, + 0.0038620668929070234, + 0.004512130748480558, + 0.00565575435757637, + 0.005493447184562683, + 0.005576475523412228, + 0.0018571665277704597, + 0.003946787677705288, + 0.0039016457740217447, + 0.0021931047085672617, + 0.0031445163767784834, + 0.0024445594754070044, + 0.00344208930619061, + 0.004028114024549723, + 0.003792797215282917, + 0.0036354295443743467, + 0.0058728470467031, + 0.004851146135479212, + 0.004497624468058348, + 0.005019643343985081, + 0.007775735575705767, + 0.006452919449657202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0007198841194622219 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015591263771057 + ], + "opposite_logprob_tokens": [ + -1.701559066772461 + ], + "logprob_reward": -0.2015591263771057, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015591263771057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16733822226524353, + 0.10023997724056244, + 0.16328568756580353, + 0.05166366323828697, + 0.10003755241632462, + 0.09468039125204086, + 0.05971037596464157, + 0.12900777161121368, + 0.0911276787519455, + 0.04290865734219551 + ], + "credit_map": [ + 0.012722664512693882, + 0.015183709561824799, + 0.030218537896871567, + 0.018511705100536346, + 0.017658624798059464, + 0.024936748668551445, + 0.022349514067173004, + 0.02575671300292015, + 0.0066412934102118015, + 0.011765644885599613, + 0.015746453776955605, + 0.015557197853922844, + 0.01174659002572298, + 0.013396719470620155, + 0.012347693555057049, + 0.01303838286548853, + 0.012034563347697258, + 0.02079269289970398, + 0.017978589981794357, + 0.016626717522740364, + 0.013783501461148262, + 0.013212821446359158, + 0.045097384601831436, + 0.023759426549077034, + 0.004275898449122906, + 0.004515916109085083, + 0.007061306852847338, + 0.004463437013328075, + 0.005609516985714436, + 0.010398960672318935, + 0.008753662928938866, + 0.006584967486560345, + 0.007640274707227945, + 0.011519085615873337, + 0.016945665702223778, + 0.006746520288288593, + 0.010586179792881012, + 0.015121853910386562, + 0.014942667447030544, + 0.016535306349396706, + 0.006066573318094015, + 0.007638975512236357, + 0.016218973323702812, + 0.008683468215167522, + 0.010063234716653824, + 0.01363252755254507, + 0.016060585156083107, + 0.01631605625152588, + 0.004889531526714563, + 0.007852857001125813, + 0.012677786871790886, + 0.005716315936297178, + 0.007939196191728115, + 0.006663538049906492, + 0.003875075839459896, + 0.010096069425344467, + 0.010193219408392906, + 0.013291611336171627, + 0.021126940846443176, + 0.0188587699085474, + 0.008688502945005894, + 0.012255962006747723, + 0.025394344702363014, + 0.019198428839445114, + 0.008145279251039028, + 0.00892118364572525, + 0.013095971196889877, + 0.007523793261498213, + 0.010138112120330334, + 0.011979483999311924, + 0.015984313562512398, + 0.015339542180299759, + 0.004353523254394531, + 0.004394549876451492, + 0.0049265301786363125, + 0.003727901494130492, + 0.004288841038942337, + 0.006077979225665331, + 0.008423611521720886, + 0.006715722382068634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.2015591263771057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075095273554325 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011075095273554325, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075095273554325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1317582130432129, + 0.07898484170436859, + 0.12469998002052307, + 0.12165215611457825, + 0.12299651652574539, + 0.10174781084060669, + 0.10133625566959381, + 0.06948502361774445, + 0.06604079902172089, + 0.08129841834306717 + ], + "credit_map": [ + 0.008481733500957489, + 0.014111331664025784, + 0.019875477999448776, + 0.014219074510037899, + 0.012272415682673454, + 0.017183620482683182, + 0.016320113092660904, + 0.02929442934691906, + 0.0064322915859520435, + 0.008009037934243679, + 0.013238392770290375, + 0.010757277719676495, + 0.009411503560841084, + 0.011177543550729752, + 0.007421841844916344, + 0.012536951340734959, + 0.008540377020835876, + 0.016196314245462418, + 0.011515336111187935, + 0.012100021354854107, + 0.010652215220034122, + 0.010820534080266953, + 0.035903364419937134, + 0.01897181011736393, + 0.006247171200811863, + 0.01789809763431549, + 0.013150284998118877, + 0.011976512148976326, + 0.0086637232452631, + 0.010652030818164349, + 0.03554116562008858, + 0.017523158341646194, + 0.008133319206535816, + 0.018216216936707497, + 0.017087554559111595, + 0.011694914661347866, + 0.015530486591160297, + 0.0188029445707798, + 0.01694304496049881, + 0.016588030382990837, + 0.006999367382377386, + 0.01189406868070364, + 0.015086114406585693, + 0.011889585293829441, + 0.008391675539314747, + 0.008269503712654114, + 0.02110409550368786, + 0.01811339519917965, + 0.011508147232234478, + 0.01658368669450283, + 0.015563980676233768, + 0.012350283563137054, + 0.010401890613138676, + 0.010129322297871113, + 0.009556258097290993, + 0.015242671594023705, + 0.006423909217119217, + 0.006972277071326971, + 0.010229579173028469, + 0.009429932571947575, + 0.006670114118605852, + 0.011249439790844917, + 0.009031820110976696, + 0.009477944113314152, + 0.004795575048774481, + 0.010156549513339996, + 0.009437448345124722, + 0.007553983014076948, + 0.009133707731962204, + 0.007205752655863762, + 0.007864797487854958, + 0.009892976842820644, + 0.005119543056935072, + 0.011220226064324379, + 0.009761621244251728, + 0.006371046416461468, + 0.013262364082038403, + 0.010024575516581535, + 0.01279693003743887, + 0.012742096558213234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.011075095273554325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping drone shot, from a bird's-eye view, slowly follows an ancient wall of weathered gray stone. It winds its way along a rugged ridge, stretching endlessly into the distance like the back of a giant dragon, until it disappears into the mist at the end of the sight. In the frame, a tiny figure, wearing a crimson windbreaker and carrying a bulging black backpack, advances steadily and rhythmically along the narrow, uneven stone path atop the wall. In the twilight hours, slanting golden sunlight bathes the scene in a warm, soft glow, casting the deep green mountains and the figures' long shadows across the ancient wall. The entire scene, set against the backdrop of this epic landscape, evokes a tranquil, contemplative atmosphere, imbued with the loneliness of personal exploration and the weight of history.", + "seed": 16170, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_13/index_415_seed_16170_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03135205690050498, + "vlm_logprob": -0.03135205690050498, + "vlm_yes_no_margin": 4.9249999761581424, + "vlm_token_credit": -0.03135205690050498, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008874262683093548 + ], + "opposite_logprob_tokens": [ + -5.633874416351318 + ], + "logprob_reward": -0.008874262683093548, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.008874262683093548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.288206547498703, + 0.0467158704996109, + 0.05855817720293999, + 0.08720015734434128, + 0.06252167373895645, + 0.03889021277427673, + 0.08765409886837006, + 0.03046499937772751, + 0.11350551247596741, + 0.1862827092409134 + ], + "credit_map": [ + 0.014304420910775661, + 0.019016733393073082, + 0.09133891761302948, + 0.08320782333612442, + 0.016999851912260056, + 0.01621018536388874, + 0.026999643072485924, + 0.020128987729549408, + 0.004728148225694895, + 0.005931603256613016, + 0.00862334854900837, + 0.008166230283677578, + 0.004707056097686291, + 0.004343452397733927, + 0.006314902100712061, + 0.003901130985468626, + 0.0062358081340789795, + 0.009835861623287201, + 0.010287400335073471, + 0.013502180576324463, + 0.003772320691496134, + 0.006567665375769138, + 0.0048021310940384865, + 0.003554808907210827, + 0.010771885514259338, + 0.012685883790254593, + 0.013478294014930725, + 0.013581146486103535, + 0.00561467744410038, + 0.011712059378623962, + 0.011623328551650047, + 0.0077328854240477085, + 0.00789723452180624, + 0.00623134383931756, + 0.010341081768274307, + 0.011302709579467773, + 0.004450798965990543, + 0.011141505092382431, + 0.005588044412434101, + 0.005568956024944782, + 0.004467005841434002, + 0.0036423460114747286, + 0.006080352235585451, + 0.007348736748099327, + 0.0027411640621721745, + 0.00720100337639451, + 0.004413999617099762, + 0.0029956044163554907, + 0.013576207682490349, + 0.0083782859146595, + 0.013023649342358112, + 0.01021308172494173, + 0.008646121248602867, + 0.015865515917539597, + 0.009704412892460823, + 0.008246822282671928, + 0.002793129999190569, + 0.003953089937567711, + 0.004017171449959278, + 0.003903739620000124, + 0.003893049666658044, + 0.004711917135864496, + 0.0040902430191636086, + 0.0031026590149849653, + 0.013841203413903713, + 0.008276615291833878, + 0.021833155304193497, + 0.011407527141273022, + 0.011586518958210945, + 0.018711315467953682, + 0.01179226953536272, + 0.016056904569268227, + 0.019029857590794563, + 0.006976613774895668, + 0.05181435868144035, + 0.01192311942577362, + 0.0170943271368742, + 0.039098773151636124, + 0.008103763684630394, + 0.032241903245449066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.008874262683093548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709005177021027 + ], + "opposite_logprob_tokens": [ + -2.1270899772644043 + ], + "logprob_reward": -0.12709005177021027, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12709005177021027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3394392132759094, + 0.05813535675406456, + 0.058303896337747574, + 0.08096976578235626, + 0.09294980764389038, + 0.06012970209121704, + 0.0809670090675354, + 0.07517091929912567, + 0.06687358766794205, + 0.08706069737672806 + ], + "credit_map": [ + 0.017028504982590675, + 0.02581881359219551, + 0.10525032132863998, + 0.08975932002067566, + 0.024366920813918114, + 0.02956480160355568, + 0.028103161603212357, + 0.019547397270798683, + 0.0051174163818359375, + 0.00901800673455, + 0.012249485589563847, + 0.010057508945465088, + 0.005197296850383282, + 0.004357044585049152, + 0.007134641986340284, + 0.005003954283893108, + 0.0059878043830394745, + 0.011922054924070835, + 0.008411356247961521, + 0.010131808929145336, + 0.005289002321660519, + 0.005929091013967991, + 0.006735223811119795, + 0.0038975581992417574, + 0.00828439462929964, + 0.009952453896403313, + 0.012356696650385857, + 0.015857793390750885, + 0.007197225466370583, + 0.008649729192256927, + 0.010893939062952995, + 0.007777536287903786, + 0.009616954252123833, + 0.007759743835777044, + 0.016227087005972862, + 0.014816041104495525, + 0.007923681288957596, + 0.01859886385500431, + 0.011091204360127449, + 0.0069162361323833466, + 0.006760737393051386, + 0.008763657882809639, + 0.010017494671046734, + 0.010209977626800537, + 0.004546275828033686, + 0.00958260428160429, + 0.005817770957946777, + 0.0044311825186014175, + 0.01118922233581543, + 0.006949790753424168, + 0.012224552221596241, + 0.009028121829032898, + 0.007784961257129908, + 0.014207194559276104, + 0.01127632986754179, + 0.008306835778057575, + 0.009867454878985882, + 0.006197523325681686, + 0.011280141770839691, + 0.010171732865273952, + 0.00814737007021904, + 0.010313019156455994, + 0.01197310071438551, + 0.007220577914267778, + 0.007541253697127104, + 0.004668030887842178, + 0.010284111835062504, + 0.010179373435676098, + 0.007690533064305782, + 0.010406982153654099, + 0.007881365716457367, + 0.008221938274800777, + 0.005692410748451948, + 0.005634342785924673, + 0.01366579718887806, + 0.008049220778048038, + 0.011980115436017513, + 0.010284441523253918, + 0.00978761911392212, + 0.02196674607694149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.12709005177021027 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006337541271932423 + ], + "opposite_logprob_tokens": [ + -7.375633716583252 + ], + "logprob_reward": -0.0006337541271932423, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006337541271932423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3138475716114044, + 0.1002303883433342, + 0.05267612263560295, + 0.10542426258325577, + 0.11024647206068039, + 0.06196969375014305, + 0.026161668822169304, + 0.04583331197500229, + 0.03588242083787918, + 0.1477281153202057 + ], + "credit_map": [ + 0.02419251948595047, + 0.05033068358898163, + 0.07811997830867767, + 0.06875143945217133, + 0.033703047782182693, + 0.013767292723059654, + 0.03195931017398834, + 0.013023266568779945, + 0.0071877711452543736, + 0.01415781769901514, + 0.01872766949236393, + 0.02272874116897583, + 0.008168281055986881, + 0.008752066642045975, + 0.01324232667684555, + 0.007265704683959484, + 0.005547417793422937, + 0.008479529060423374, + 0.008260669186711311, + 0.010777709074318409, + 0.004359350074082613, + 0.0050008296966552734, + 0.005894949194043875, + 0.004355665296316147, + 0.011470668017864227, + 0.022770967334508896, + 0.015594685450196266, + 0.014292754232883453, + 0.007324900012463331, + 0.014728688634932041, + 0.009704291820526123, + 0.0095372898504138, + 0.015766160562634468, + 0.009929084219038486, + 0.018325287848711014, + 0.01868361420929432, + 0.00848356168717146, + 0.016159100458025932, + 0.01277962513267994, + 0.01012002769857645, + 0.007176084443926811, + 0.006775106303393841, + 0.01044976245611906, + 0.012992373667657375, + 0.004972767550498247, + 0.011506921611726284, + 0.004157009534537792, + 0.003939665853977203, + 0.003532337723299861, + 0.002751067513599992, + 0.004464587662369013, + 0.005316806025803089, + 0.0024231651332229376, + 0.003503599902614951, + 0.0022913541179150343, + 0.0018787465523928404, + 0.0060845329426229, + 0.004390890244394541, + 0.00670997891575098, + 0.00741609325632453, + 0.005903278477489948, + 0.006077861879020929, + 0.005342094227671623, + 0.003908579703420401, + 0.0037657024804502726, + 0.002638204488903284, + 0.005537816323339939, + 0.005641091149300337, + 0.003770922776311636, + 0.007387787103652954, + 0.003990813158452511, + 0.003150082193315029, + 0.015221458859741688, + 0.009923089295625687, + 0.02140895463526249, + 0.010801032185554504, + 0.03045361302793026, + 0.02366694062948227, + 0.012710357084870338, + 0.023542659357190132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0006337541271932423 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067545861005783 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.016067545861005783, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067545861005783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27225208282470703, + 0.05644353851675987, + 0.05122974142432213, + 0.0675249844789505, + 0.3050116002559662, + 0.035054437816143036, + 0.023086482658982277, + 0.010646462440490723, + 0.03406723961234093, + 0.14468346536159515 + ], + "credit_map": [ + 0.019221104681491852, + 0.034890253096818924, + 0.08343067020177841, + 0.07403859496116638, + 0.021857978776097298, + 0.008770132437348366, + 0.02156500332057476, + 0.00847832951694727, + 0.005211888812482357, + 0.007335298229008913, + 0.011404868215322495, + 0.012504016980528831, + 0.004864951595664024, + 0.005347646307200193, + 0.005599117372184992, + 0.0041757505387067795, + 0.0057759592309594154, + 0.01195978932082653, + 0.008153064176440239, + 0.009250398725271225, + 0.004107050132006407, + 0.0041206637397408485, + 0.004356293007731438, + 0.0035065244883298874, + 0.009794988669455051, + 0.010905827395617962, + 0.009894008748233318, + 0.009672688320279121, + 0.004019560292363167, + 0.009237680584192276, + 0.007598435506224632, + 0.006401794031262398, + 0.04440046101808548, + 0.02986030839383602, + 0.049658045172691345, + 0.056381870061159134, + 0.02103826403617859, + 0.04903382062911987, + 0.03269539400935173, + 0.021943438798189163, + 0.005544030107557774, + 0.00402766652405262, + 0.00630949018523097, + 0.005165197886526585, + 0.0029086871072649956, + 0.005663459654897451, + 0.002996037947013974, + 0.0024398674722760916, + 0.0035329621750861406, + 0.0020911472383886576, + 0.0033446659799665213, + 0.0032075338531285524, + 0.002199997426941991, + 0.004487817641347647, + 0.0021986982319504023, + 0.0020236598793417215, + 0.0011146307224407792, + 0.0008519142866134644, + 0.002116756746545434, + 0.001314207329414785, + 0.001340968650765717, + 0.00147172249853611, + 0.0013932486763224006, + 0.0010430129477754235, + 0.0037580253556370735, + 0.0024947435595095158, + 0.004695920739322901, + 0.007754020858556032, + 0.0032916173804551363, + 0.004112652037292719, + 0.004596438724547625, + 0.003363820258527994, + 0.014617322012782097, + 0.00481291813775897, + 0.036055196076631546, + 0.005309459287673235, + 0.015041415579617023, + 0.04618661478161812, + 0.006153373513370752, + 0.016507159918546677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.016067545861005783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping bird's-eye view following an ancient wall of weathered gray stone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a tiny figure wearing a crimson windbreaker and carrying a bulging black backpack visible advancing along the wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a rugged ridge with deep green mountains stretching into the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by slanting golden sunlight in twilight hours casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly follow the wall as it winds along the ridge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094670061022043 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094670061022043, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094670061022043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09211952984333038, + 0.0741557776927948, + 0.04118209332227707, + 0.0467681922018528, + 0.06936829537153244, + 0.058358367532491684, + 0.03571121022105217, + 0.11342530697584152, + 0.09120532870292664, + 0.3777058720588684 + ], + "credit_map": [ + 0.006388802547007799, + 0.01021442748606205, + 0.024395134299993515, + 0.018713127821683884, + 0.00865116436034441, + 0.00693588936701417, + 0.009360017254948616, + 0.0074609690345823765, + 0.006169252563267946, + 0.013683446682989597, + 0.012776011601090431, + 0.013954452238976955, + 0.005618392489850521, + 0.005914293695241213, + 0.008693577721714973, + 0.0073463572189211845, + 0.005050841253250837, + 0.009293017908930779, + 0.006571587175130844, + 0.006384999956935644, + 0.0026148438919335604, + 0.005203888285905123, + 0.0032673985697329044, + 0.0027955188415944576, + 0.004904548171907663, + 0.006547039374709129, + 0.007769504562020302, + 0.00891860667616129, + 0.004288593772798777, + 0.0038752604741603136, + 0.005478194449096918, + 0.004986452870070934, + 0.0078141363337636, + 0.006364472676068544, + 0.013773739337921143, + 0.01599138416349888, + 0.0042607132345438, + 0.009494167752563953, + 0.007020091637969017, + 0.004649593029171228, + 0.009296047501266003, + 0.008262412622570992, + 0.0092126140370965, + 0.0066047185100615025, + 0.004123827908188105, + 0.012686779722571373, + 0.0046151443384587765, + 0.0035568242892622948, + 0.003509301459416747, + 0.003410882316529751, + 0.006051965523511171, + 0.006211346946656704, + 0.003560093929991126, + 0.004201628267765045, + 0.0048453896306455135, + 0.0039206030778586864, + 0.014667373150587082, + 0.011996045708656311, + 0.02246803604066372, + 0.015006507746875286, + 0.013893726281821728, + 0.014208623208105564, + 0.00953887403011322, + 0.011646122671663761, + 0.010160470381379128, + 0.006237394642084837, + 0.017971783876419067, + 0.011279973201453686, + 0.009190157055854797, + 0.012514948844909668, + 0.011027757078409195, + 0.012822847813367844, + 0.0356827937066555, + 0.012691874988377094, + 0.10342058539390564, + 0.021147064864635468, + 0.030488502234220505, + 0.10925977677106857, + 0.013732878491282463, + 0.05128239467740059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.004094670061022043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_14/index_tensor([1461, 1461])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_14/index_tensor([1461, 1461])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f905e85b28fef11d06fbfbadb866b22967242862 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_14/index_tensor([1461, 1461])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide-angle lens captures the scene from above the seafloor, the camera drifting to the right in unusually slow and steady lateral movements, creating a sense of reverent observation. In the profound azure silence of the deep ocean, a massive great white shark rests, hovering almost motionless above a pale, fine-sand seafloor dotted with gentle ripples from the current. Its massive body, a masterpiece of predatory evolution, is rendered in surreal detail; the slate-gray skin of its dorsal side, delicately marked with faint silver scars, contrasts starkly with the stark, clean white of its underbelly. Its black, hollow eyes gaze forward into the abyss, while its jaw, slightly ajar, offers a glimpse of jagged, triangular teeth. The only perceptible movement is the slow, rhythmic opening of its gill slits with breathing and the gentle, almost lazy drift of its powerful tail fin on the invisible current. Bright, ethereal sunlight penetrates the water from the distant surface, creating unique \"god rays\" that pierce the deep blue and illuminate swirling, dust-like particles of marine snow. These rays punctuate the shark’s powerful back, casting an ever-changing mosaic of light and shadow, highlighting its muscular silhouette and creating an atmosphere of tranquility yet tension.", + "seed": 26422, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_14/index_1461_seed_26422_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.34218199234455826, + "vlm_logprob": -0.34218199234455826, + "vlm_yes_no_margin": 2.0, + "vlm_token_credit": -0.34218199234455826, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8434120416641235 + ], + "opposite_logprob_tokens": [ + -0.5934120416641235 + ], + "logprob_reward": -0.8434120416641235, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8434120416641235, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22280685603618622, + 0.07339596003293991, + 0.06424295157194138, + 0.12884484231472015, + 0.09509547799825668, + 0.054670654237270355, + 0.09063594788312912, + 0.11925914883613586, + 0.07829834520816803, + 0.0727498009800911 + ], + "credit_map": [ + 0.021252065896987915, + 0.016041304916143417, + 0.042054638266563416, + 0.02525443397462368, + 0.012510471045970917, + 0.016914498060941696, + 0.05195159465074539, + 0.03682783991098404, + 0.010556076653301716, + 0.006691355723887682, + 0.014027412980794907, + 0.008786524645984173, + 0.0054354420863091946, + 0.005167281720787287, + 0.010737505741417408, + 0.011994368396699429, + 0.009518102742731571, + 0.005922995507717133, + 0.011164610274136066, + 0.008794523775577545, + 0.004717053379863501, + 0.005223867949098349, + 0.010214725509285927, + 0.008687077090144157, + 0.0159597247838974, + 0.011284943670034409, + 0.022150477394461632, + 0.014735447242856026, + 0.010875916108489037, + 0.012925436720252037, + 0.021075420081615448, + 0.019837478175759315, + 0.013660436496138573, + 0.007251758128404617, + 0.01597015932202339, + 0.012823679484426975, + 0.00843567680567503, + 0.008997263386845589, + 0.014709645882248878, + 0.013246862217783928, + 0.007629316300153732, + 0.005069427657872438, + 0.008406508713960648, + 0.005855761468410492, + 0.006248597055673599, + 0.006068149581551552, + 0.007773152086883783, + 0.0076197427697479725, + 0.010806034319102764, + 0.010633128695189953, + 0.012332886457443237, + 0.009393622167408466, + 0.0134584316983819, + 0.005999491084367037, + 0.016450339928269386, + 0.011562017723917961, + 0.0094175199046731, + 0.01192228589206934, + 0.021324513480067253, + 0.010283004492521286, + 0.011801555752754211, + 0.010375868529081345, + 0.027130529284477234, + 0.017003880813717842, + 0.0060106441378593445, + 0.009890207089483738, + 0.00993510615080595, + 0.01004442572593689, + 0.008901780471205711, + 0.007656115107238293, + 0.015913423150777817, + 0.009946652688086033, + 0.004668916575610638, + 0.007520709186792374, + 0.00766099663451314, + 0.014946048147976398, + 0.007616604678332806, + 0.007159344851970673, + 0.01184872817248106, + 0.011328455060720444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.8434120416641235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328380703926086 + ], + "opposite_logprob_tokens": [ + -0.7578380703926086 + ], + "logprob_reward": -0.6328380703926086, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328380703926086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18663723766803741, + 0.08895322680473328, + 0.053659968078136444, + 0.07577567547559738, + 0.07773909717798233, + 0.04996629059314728, + 0.05559232831001282, + 0.19787102937698364, + 0.12971557676792145, + 0.08408956974744797 + ], + "credit_map": [ + 0.027911262586712837, + 0.015953993424773216, + 0.044466305524110794, + 0.017241552472114563, + 0.013877774588763714, + 0.01776180788874626, + 0.02027474343776703, + 0.029149791225790977, + 0.012273711152374744, + 0.010626381263136864, + 0.013919073157012463, + 0.011941474862396717, + 0.005726929288357496, + 0.007825550623238087, + 0.013256183825433254, + 0.01338391937315464, + 0.0064901262521743774, + 0.0067980531603097916, + 0.009970191866159439, + 0.006198250688612461, + 0.004506631754338741, + 0.00438833748921752, + 0.007296342868357897, + 0.008012032136321068, + 0.009291324764490128, + 0.01026744395494461, + 0.010762147605419159, + 0.009519570507109165, + 0.009527644142508507, + 0.006254401057958603, + 0.010124187916517258, + 0.010028956457972527, + 0.008544503711163998, + 0.006426162086427212, + 0.012667986564338207, + 0.013244596309959888, + 0.007304786238819361, + 0.007302739657461643, + 0.012373670935630798, + 0.00987465400248766, + 0.006013273261487484, + 0.005427104886621237, + 0.008825711905956268, + 0.005199295002967119, + 0.004910048097372055, + 0.005179164931178093, + 0.006837960332632065, + 0.007573735900223255, + 0.006741379853338003, + 0.0051651312969625, + 0.009173060767352581, + 0.0066508580930531025, + 0.006079958286136389, + 0.006820916198194027, + 0.007155774626880884, + 0.007805244997143745, + 0.016189852729439735, + 0.021347664296627045, + 0.031072640791535378, + 0.018228527158498764, + 0.01767737790942192, + 0.01752430573105812, + 0.04781733825802803, + 0.028013326227664948, + 0.008856031112372875, + 0.015358131378889084, + 0.014365611597895622, + 0.018291866406798363, + 0.021454455330967903, + 0.012591943144798279, + 0.022175095975399017, + 0.01662244275212288, + 0.006602136883884668, + 0.011175872758030891, + 0.012377743609249592, + 0.0111840246245265, + 0.0104116415604949, + 0.009418994188308716, + 0.011259635910391808, + 0.011659511364996433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.6328380703926086 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006753121502697468 + ], + "opposite_logprob_tokens": [ + -5.006752967834473 + ], + "logprob_reward": -0.006753121502697468, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006753121502697468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.266103595495224, + 0.25780245661735535, + 0.05380222946405411, + 0.05103734880685806, + 0.0940825417637825, + 0.038578666746616364, + 0.054064225405454636, + 0.05921444296836853, + 0.05502939224243164, + 0.07028511166572571 + ], + "credit_map": [ + 0.024468736723065376, + 0.02294454164803028, + 0.04297101870179176, + 0.03358114883303642, + 0.01720830611884594, + 0.020465563982725143, + 0.06667326390743256, + 0.03779101371765137, + 0.02071186900138855, + 0.03147954121232033, + 0.043460629880428314, + 0.04374929890036583, + 0.019645262509584427, + 0.022083166986703873, + 0.0381220243871212, + 0.03855067491531372, + 0.00735159358009696, + 0.005483292043209076, + 0.009967253543436527, + 0.007476831786334515, + 0.004127323627471924, + 0.003165888600051403, + 0.009618044830858707, + 0.006611996795982122, + 0.006258963141590357, + 0.0048332600854337215, + 0.009778104722499847, + 0.005672717001289129, + 0.0038856484461575747, + 0.005267309490591288, + 0.007571275345981121, + 0.007770067546516657, + 0.009658150374889374, + 0.007297116331756115, + 0.01779293082654476, + 0.013858967460691929, + 0.008908331394195557, + 0.007716526743024588, + 0.013907019980251789, + 0.01494350004941225, + 0.003600820666179061, + 0.0036398256197571754, + 0.007487693335860968, + 0.004980653524398804, + 0.0034041767939925194, + 0.004400473088026047, + 0.005864238832145929, + 0.005200786050409079, + 0.005386385601013899, + 0.00513258995488286, + 0.0072774034924805164, + 0.006487421691417694, + 0.007051149848848581, + 0.005244717001914978, + 0.010877943597733974, + 0.006606612354516983, + 0.0038905302062630653, + 0.005347508005797863, + 0.007006489206105471, + 0.008053379133343697, + 0.012369515374302864, + 0.007630996871739626, + 0.006135168019682169, + 0.008780858479440212, + 0.0032962101977318525, + 0.005166263785213232, + 0.007172086276113987, + 0.007120422553271055, + 0.007701790891587734, + 0.004704904276877642, + 0.012040364556014538, + 0.007827351801097393, + 0.003922170028090477, + 0.005964634474366903, + 0.008227542042732239, + 0.014659604988992214, + 0.008700569160282612, + 0.008097226731479168, + 0.009350950829684734, + 0.01136241015046835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.006753121502697468 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354903355240822 + ], + "opposite_logprob_tokens": [ + -3.6513547897338867 + ], + "logprob_reward": -0.026354903355240822, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354903355240822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13740752637386322, + 0.20801115036010742, + 0.2230023741722107, + 0.04400724172592163, + 0.14738057553768158, + 0.029755720868706703, + 0.04953906685113907, + 0.043019555509090424, + 0.07467744499444962, + 0.043199412524700165 + ], + "credit_map": [ + 0.01439851988106966, + 0.013361412100493908, + 0.027957478538155556, + 0.015392530709505081, + 0.009410898201167583, + 0.012426271103322506, + 0.02623617649078369, + 0.018224231898784637, + 0.02389317750930786, + 0.017680544406175613, + 0.05399075150489807, + 0.02426297776401043, + 0.014274483546614647, + 0.014117075130343437, + 0.02482912316918373, + 0.03496300056576729, + 0.034766390919685364, + 0.01914549060165882, + 0.032892972230911255, + 0.02746797353029251, + 0.015457565896213055, + 0.012185195460915565, + 0.0394011028110981, + 0.04168563708662987, + 0.00481799291446805, + 0.004044860601425171, + 0.007790928706526756, + 0.0060123600997030735, + 0.0033268267288804054, + 0.0042802379466593266, + 0.007539024110883474, + 0.006195004563778639, + 0.02189505286514759, + 0.01741756498813629, + 0.021262114867568016, + 0.020633533596992493, + 0.015720898285508156, + 0.01030865591019392, + 0.018529271706938744, + 0.0216134674847126, + 0.003294010879471898, + 0.002443834440782666, + 0.0035487753339111805, + 0.004335897509008646, + 0.002316097728908062, + 0.0025885889772325754, + 0.00583738973364234, + 0.005391123238950968, + 0.005733969621360302, + 0.005278095602989197, + 0.007033356465399265, + 0.005769426468759775, + 0.005098856054246426, + 0.00403282605111599, + 0.010220082476735115, + 0.006372447125613689, + 0.004570960998535156, + 0.007323798257857561, + 0.005659632850438356, + 0.0037306835874915123, + 0.004377740900963545, + 0.003525608219206333, + 0.008733632043004036, + 0.005097491666674614, + 0.004520897287875414, + 0.005873383954167366, + 0.010368824005126953, + 0.010843528434634209, + 0.00909250695258379, + 0.008329031057655811, + 0.017543362453579903, + 0.008105902932584286, + 0.003025621874257922, + 0.0028328686021268368, + 0.0069991303607821465, + 0.007868854328989983, + 0.0037115721497684717, + 0.003060573246330023, + 0.008090018294751644, + 0.007610768545418978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.026354903355240822 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155182480812073 + ], + "opposite_logprob_tokens": [ + -1.7015517950057983 + ], + "logprob_reward": -0.20155182480812073, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155182480812073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14296133816242218, + 0.08357512950897217, + 0.09547694027423859, + 0.10502734035253525, + 0.1229131817817688, + 0.13287454843521118, + 0.047357358038425446, + 0.10375793278217316, + 0.0721752718091011, + 0.09388097375631332 + ], + "credit_map": [ + 0.013208089396357536, + 0.012681537307798862, + 0.023985950276255608, + 0.017237281426787376, + 0.00860463734716177, + 0.01146178413182497, + 0.03565782681107521, + 0.02012423425912857, + 0.00907042808830738, + 0.011481652967631817, + 0.015930715948343277, + 0.010379417799413204, + 0.006252234801650047, + 0.004712732508778572, + 0.01067147497087717, + 0.015076465904712677, + 0.009178035892546177, + 0.007569512352347374, + 0.017535395920276642, + 0.015829524025321007, + 0.00602224376052618, + 0.010372094810009003, + 0.015089185908436775, + 0.013880941085517406, + 0.00793380569666624, + 0.012154917232692242, + 0.01749158836901188, + 0.015350086614489555, + 0.00893201120197773, + 0.012481598183512688, + 0.015344761312007904, + 0.015338570810854435, + 0.012770813889801502, + 0.011222796514630318, + 0.0171526987105608, + 0.02129131741821766, + 0.012372412718832493, + 0.011694706045091152, + 0.01693195104598999, + 0.01947648450732231, + 0.014540812000632286, + 0.020634131506085396, + 0.020914249122142792, + 0.018309839069843292, + 0.01278003491461277, + 0.01317992340773344, + 0.013506756164133549, + 0.01900879479944706, + 0.004522935952991247, + 0.004890024662017822, + 0.006010495591908693, + 0.005763696040958166, + 0.004473743494600058, + 0.003486155066639185, + 0.011998139321804047, + 0.00621216744184494, + 0.007819653488695621, + 0.01323341391980648, + 0.00952964834868908, + 0.014254975132644176, + 0.01096602063626051, + 0.007718333974480629, + 0.026736032217741013, + 0.013499859720468521, + 0.005332574248313904, + 0.0068020448088645935, + 0.01058861892670393, + 0.008978825062513351, + 0.009682872332632542, + 0.007959690876305103, + 0.012639554217457771, + 0.010191094130277634, + 0.0067025888711214066, + 0.007029676344245672, + 0.013241305947303772, + 0.017982248216867447, + 0.01014696341007948, + 0.012193010188639164, + 0.013466974720358849, + 0.013118207454681396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.20155182480812073 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle lens captures the scene from above the seafloor, the camera drifting to the right in unusually slow and steady lateral movements, creating a sense of reverent observation. In the profound azure silence of the deep ocean, a massive great white shark rests, hovering almost motionless above a pale, fine-sand seafloor dotted with gentle ripples from the current. Its massive body, a masterpiece of predatory evolution, is rendered in surreal detail; the slate-gray skin of its dorsal side, delicately marked with faint silver scars, contrasts starkly with the stark, clean white of its underbelly. Its black, hollow eyes gaze forward into the abyss, while its jaw, slightly ajar, offers a glimpse of jagged, triangular teeth. The only perceptible movement is the slow, rhythmic opening of its gill slits with breathing and the gentle, almost lazy drift of its powerful tail fin on the invisible current. Bright, ethereal sunlight penetrates the water from the distant surface, creating unique \"god rays\" that pierce the deep blue and illuminate swirling, dust-like particles of marine snow. These rays punctuate the shark’s powerful back, casting an ever-changing mosaic of light and shadow, highlighting its muscular silhouette and creating an atmosphere of tranquility yet tension.", + "seed": 26423, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_14/index_1461_seed_26423_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1139225685968995, + "vlm_logprob": -0.1139225685968995, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.1139225685968995, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.36981165409088135 + ], + "opposite_logprob_tokens": [ + -1.2448116540908813 + ], + "logprob_reward": -0.36981165409088135, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.36981165409088135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1518033742904663, + 0.09463822841644287, + 0.10527749359607697, + 0.10028679668903351, + 0.22704678773880005, + 0.11494043469429016, + 0.0867476537823677, + 0.04296917840838432, + 0.048895932734012604, + 0.02739410474896431 + ], + "credit_map": [ + 0.017137913033366203, + 0.022041065618395805, + 0.02788517065346241, + 0.033949606120586395, + 0.014562828466296196, + 0.009560316801071167, + 0.013484862633049488, + 0.01318160817027092, + 0.021989675238728523, + 0.018930107355117798, + 0.01508261077105999, + 0.020311987027525902, + 0.004459868185222149, + 0.003280295291915536, + 0.006170997396111488, + 0.0044126869179308414, + 0.015035429038107395, + 0.01242949441075325, + 0.014429103583097458, + 0.020865101367235184, + 0.009272287599742413, + 0.008547364734113216, + 0.012969247996807098, + 0.011729465797543526, + 0.01688435673713684, + 0.008028938435018063, + 0.019006559625267982, + 0.012433446943759918, + 0.01074750442057848, + 0.010837804526090622, + 0.009181700646877289, + 0.013166479766368866, + 0.04065521061420441, + 0.020038487389683723, + 0.05587676167488098, + 0.02040988579392433, + 0.021116551011800766, + 0.02746671810746193, + 0.018717439845204353, + 0.022765744477510452, + 0.018397822976112366, + 0.006969572510570288, + 0.016993775963783264, + 0.009678366594016552, + 0.008769485168159008, + 0.030046336352825165, + 0.010129064321517944, + 0.013956008478999138, + 0.007606918457895517, + 0.006312629207968712, + 0.027673708274960518, + 0.010362537577748299, + 0.005848687142133713, + 0.013628005981445312, + 0.00922638364136219, + 0.006088779307901859, + 0.004503637086600065, + 0.005253396928310394, + 0.006774332839995623, + 0.005505379755049944, + 0.003743964247405529, + 0.005135766696184874, + 0.008494866080582142, + 0.0035578333772718906, + 0.00582132488489151, + 0.004085861146450043, + 0.010585932061076164, + 0.005639286246150732, + 0.004703192040324211, + 0.005408231168985367, + 0.00688707223162055, + 0.0057650343514978886, + 0.0025969105772674084, + 0.0019071957794949412, + 0.0029592965729534626, + 0.004245053045451641, + 0.0022866884246468544, + 0.00512102572247386, + 0.004202940501272678, + 0.00407499261200428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.36981165409088135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812139853835106 + ], + "opposite_logprob_tokens": [ + -3.288121461868286 + ], + "logprob_reward": -0.03812139853835106, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812139853835106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1890150010585785, + 0.09826045483350754, + 0.19752267003059387, + 0.24017538130283356, + 0.06249179318547249, + 0.07916467636823654, + 0.03396110236644745, + 0.04152597114443779, + 0.023103727027773857, + 0.034779202193021774 + ], + "credit_map": [ + 0.04187857732176781, + 0.04455989971756935, + 0.0309455506503582, + 0.026327969506382942, + 0.011989712715148926, + 0.009641953743994236, + 0.011358157731592655, + 0.0123131787404418, + 0.01595095731317997, + 0.023023445159196854, + 0.012921212241053581, + 0.014939217828214169, + 0.007468882016837597, + 0.00914941169321537, + 0.0077683571726083755, + 0.007038972806185484, + 0.052063532173633575, + 0.019026409834623337, + 0.052675195038318634, + 0.021437568590044975, + 0.01376430131494999, + 0.00873799342662096, + 0.016688402742147446, + 0.013129260390996933, + 0.04772182181477547, + 0.01428858283907175, + 0.06275668740272522, + 0.02401440218091011, + 0.02198643609881401, + 0.02543792501091957, + 0.01583337038755417, + 0.028136175125837326, + 0.0098951430991292, + 0.007011492270976305, + 0.011004200205206871, + 0.004874074831604958, + 0.007562172599136829, + 0.00878127384930849, + 0.005302607547491789, + 0.00806083157658577, + 0.011461067013442516, + 0.005935879424214363, + 0.010528167709708214, + 0.007507567759603262, + 0.006385712884366512, + 0.017002955079078674, + 0.010769028216600418, + 0.009574301540851593, + 0.002793255029246211, + 0.0030184760689735413, + 0.006899002008140087, + 0.0051474301144480705, + 0.0035038571804761887, + 0.0033120629377663136, + 0.0063811796717345715, + 0.002905838191509247, + 0.004795132204890251, + 0.00422386871650815, + 0.009025078266859055, + 0.0044476124458014965, + 0.00393443601205945, + 0.005879970733076334, + 0.005531023256480694, + 0.0036888481117784977, + 0.002933588344603777, + 0.0019535417668521404, + 0.003252202644944191, + 0.002552544232457876, + 0.002305813832208514, + 0.0036216378211975098, + 0.0031928454991430044, + 0.003291552187874913, + 0.0030784481205046177, + 0.003348554717376828, + 0.004833439365029335, + 0.004908131901174784, + 0.0025072230491787195, + 0.008213740773499012, + 0.0038306834176182747, + 0.004058979917317629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.03812139853835106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028257938101887703 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028257938101887703, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028257938101887703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22382666170597076, + 0.07444246858358383, + 0.13660158216953278, + 0.08078315109014511, + 0.17662915587425232, + 0.1384257972240448, + 0.05611763149499893, + 0.06700780242681503, + 0.025427531450986862, + 0.02073821984231472 + ], + "credit_map": [ + 0.026517443358898163, + 0.03246321156620979, + 0.02742348238825798, + 0.0356341116130352, + 0.030949151143431664, + 0.023245293647050858, + 0.027748972177505493, + 0.01984500326216221, + 0.014641134068369865, + 0.021292677149176598, + 0.004961432423442602, + 0.00904403068125248, + 0.010094137862324715, + 0.0037756396923214197, + 0.006500019691884518, + 0.004133398178964853, + 0.017491187900304794, + 0.018378639593720436, + 0.01477778423577547, + 0.017965145409107208, + 0.021761121228337288, + 0.009512970224022865, + 0.020717840641736984, + 0.0159969013184309, + 0.009241950698196888, + 0.008579349145293236, + 0.009809339419007301, + 0.011563141830265522, + 0.013823743909597397, + 0.007851836271584034, + 0.009072627872228622, + 0.010841156356036663, + 0.025540180504322052, + 0.01544578280299902, + 0.03700592741370201, + 0.017203398048877716, + 0.021666867658495903, + 0.020383689552545547, + 0.01698347181081772, + 0.022399835288524628, + 0.01835666596889496, + 0.006060583051294088, + 0.04365965723991394, + 0.010317385196685791, + 0.008259043097496033, + 0.023435626178979874, + 0.014933680184185505, + 0.013403166085481644, + 0.005177456885576248, + 0.007671366911381483, + 0.009254295378923416, + 0.009265976026654243, + 0.007887021638453007, + 0.006314036902040243, + 0.006215300876647234, + 0.00433217640966177, + 0.007092153653502464, + 0.008449896238744259, + 0.01349126361310482, + 0.0057001919485628605, + 0.006283382885158062, + 0.009583196602761745, + 0.009857182390987873, + 0.006550535559654236, + 0.0030109600629657507, + 0.0022680554538965225, + 0.0031413985416293144, + 0.0029395627789199352, + 0.003947332967072725, + 0.0028570385184139013, + 0.003536521689966321, + 0.0037266616709530354, + 0.0024352262262254953, + 0.0022049976978451014, + 0.0030199424363672733, + 0.002920570084825158, + 0.002363752806559205, + 0.002108935732394457, + 0.002453753724694252, + 0.0032310409005731344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0028257938101887703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067076474428177 + ], + "opposite_logprob_tokens": [ + -4.141067028045654 + ], + "logprob_reward": -0.016067076474428177, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067076474428177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2051869034767151, + 0.23533909022808075, + 0.1321406215429306, + 0.11507898569107056, + 0.08364326506853104, + 0.10644486546516418, + 0.04631384462118149, + 0.030425656586885452, + 0.032057445496320724, + 0.013369325548410416 + ], + "credit_map": [ + 0.05116289108991623, + 0.054606035351753235, + 0.028205282986164093, + 0.024799592792987823, + 0.01102455984801054, + 0.01186649315059185, + 0.010485269129276276, + 0.01303676888346672, + 0.047346796840429306, + 0.07554735243320465, + 0.024847444146871567, + 0.04748086631298065, + 0.011555149219930172, + 0.0074402703903615475, + 0.011400614865124226, + 0.009720606729388237, + 0.02504655532538891, + 0.014979842118918896, + 0.02110299840569496, + 0.01761450804769993, + 0.015314671210944653, + 0.009728309698402882, + 0.01701114885509014, + 0.011342577636241913, + 0.021607723087072372, + 0.00818172562867403, + 0.022991595789790154, + 0.010591798461973667, + 0.014407191425561905, + 0.013203511945903301, + 0.009787213988602161, + 0.014308225363492966, + 0.014469943940639496, + 0.008821352384984493, + 0.01711289770901203, + 0.007829437032341957, + 0.009733564220368862, + 0.011372575536370277, + 0.0068528964184224606, + 0.007450604811310768, + 0.016061728820204735, + 0.003504750784486532, + 0.03509410470724106, + 0.006775814574211836, + 0.004446170758455992, + 0.020794671028852463, + 0.009665659628808498, + 0.010101955384016037, + 0.004831892903894186, + 0.004896028432995081, + 0.007016135845333338, + 0.006465431302785873, + 0.004980804864317179, + 0.005633173510432243, + 0.005187997594475746, + 0.007302380632609129, + 0.004572421777993441, + 0.0033231896813958883, + 0.004940013401210308, + 0.003530119312927127, + 0.0022283007856458426, + 0.0034778406843543053, + 0.00414177356287837, + 0.004211998078972101, + 0.0038068469148129225, + 0.0027903378941118717, + 0.006300203036516905, + 0.004631598014384508, + 0.002515970030799508, + 0.0037203829269856215, + 0.0045021227560937405, + 0.0037899850867688656, + 0.0013659559190273285, + 0.001277624978683889, + 0.0022425081115216017, + 0.0013494158629328012, + 0.0013236498925834894, + 0.002604162320494652, + 0.0015958815347403288, + 0.0016101279761642218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.016067076474428177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427869200706482 + ], + "opposite_logprob_tokens": [ + -2.017786979675293 + ], + "logprob_reward": -0.1427869200706482, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1427869200706482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3805796802043915, + 0.053778231143951416, + 0.18154986202716827, + 0.1600651741027832, + 0.04696502536535263, + 0.0405203253030777, + 0.04876875504851341, + 0.036176055669784546, + 0.019312020391225815, + 0.032284948974847794 + ], + "credit_map": [ + 0.07170993089675903, + 0.0778157114982605, + 0.07759128510951996, + 0.06026773154735565, + 0.026705509051680565, + 0.022011131048202515, + 0.0224705059081316, + 0.022007880732417107, + 0.017801478505134583, + 0.018222147598862648, + 0.004558201413601637, + 0.004044502042233944, + 0.0031640774104744196, + 0.0018717811908572912, + 0.002117542317137122, + 0.0019985006656497717, + 0.05440755933523178, + 0.016510797664523125, + 0.06449806690216064, + 0.018496796488761902, + 0.007765095680952072, + 0.0057940674014389515, + 0.0068741594441235065, + 0.007203327026218176, + 0.03151881694793701, + 0.009878168813884258, + 0.04204627498984337, + 0.016366101801395416, + 0.018281137570738792, + 0.015559444203972816, + 0.010950088500976562, + 0.015465127304196358, + 0.0068995775654911995, + 0.004391215741634369, + 0.007292402908205986, + 0.003735671751201153, + 0.004672562703490257, + 0.010633506812155247, + 0.004120378755033016, + 0.005219708196818829, + 0.004861591383814812, + 0.0029001825023442507, + 0.010007110424339771, + 0.0035813134163618088, + 0.0036704076919704676, + 0.006567888893187046, + 0.004261147230863571, + 0.004670684691518545, + 0.0044973623007535934, + 0.006403551436960697, + 0.005844030529260635, + 0.009149771183729172, + 0.007376415655016899, + 0.005019017960876226, + 0.0054862769320607185, + 0.004992328118532896, + 0.003753899596631527, + 0.0029412242583930492, + 0.007211288902908564, + 0.003451880533248186, + 0.002870511030778289, + 0.007277848664671183, + 0.005490534473210573, + 0.0031788647174835205, + 0.0018484829925000668, + 0.002129933563992381, + 0.002327319234609604, + 0.002811669372022152, + 0.002342398278415203, + 0.002959228353574872, + 0.0025603934191167355, + 0.0023325947113335133, + 0.0040572467260062695, + 0.004864032845944166, + 0.002763130934908986, + 0.004343225620687008, + 0.0052104527130723, + 0.0029811635613441467, + 0.0032384407240897417, + 0.004827254451811314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.1427869200706482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle lens captures the scene from above the seafloor, the camera drifting to the right in unusually slow and steady lateral movements, creating a sense of reverent observation. In the profound azure silence of the deep ocean, a massive great white shark rests, hovering almost motionless above a pale, fine-sand seafloor dotted with gentle ripples from the current. Its massive body, a masterpiece of predatory evolution, is rendered in surreal detail; the slate-gray skin of its dorsal side, delicately marked with faint silver scars, contrasts starkly with the stark, clean white of its underbelly. Its black, hollow eyes gaze forward into the abyss, while its jaw, slightly ajar, offers a glimpse of jagged, triangular teeth. The only perceptible movement is the slow, rhythmic opening of its gill slits with breathing and the gentle, almost lazy drift of its powerful tail fin on the invisible current. Bright, ethereal sunlight penetrates the water from the distant surface, creating unique \"god rays\" that pierce the deep blue and illuminate swirling, dust-like particles of marine snow. These rays punctuate the shark’s powerful back, casting an ever-changing mosaic of light and shadow, highlighting its muscular silhouette and creating an atmosphere of tranquility yet tension.", + "seed": 26424, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_14/index_1461_seed_26424_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.29281780757009984, + "vlm_logprob": -0.29281780757009984, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.29281780757009984, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7720062732696533 + ], + "opposite_logprob_tokens": [ + -0.6470062732696533 + ], + "logprob_reward": -0.7720062732696533, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7720062732696533, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24775151908397675, + 0.09782470762729645, + 0.04167059808969498, + 0.052352722734212875, + 0.081294946372509, + 0.16627490520477295, + 0.06399044394493103, + 0.058048881590366364, + 0.13156503438949585, + 0.059226203709840775 + ], + "credit_map": [ + 0.03332710638642311, + 0.040039025247097015, + 0.03983968868851662, + 0.03308183327317238, + 0.021853933110833168, + 0.021423863247036934, + 0.02600092440843582, + 0.03218517452478409, + 0.009179169312119484, + 0.015539850108325481, + 0.01114354282617569, + 0.011545206420123577, + 0.011416235938668251, + 0.01167553011327982, + 0.01443384774029255, + 0.012891325168311596, + 0.005789718125015497, + 0.004172454122453928, + 0.008412133902311325, + 0.006916339509189129, + 0.003973797429352999, + 0.002959486562758684, + 0.004517366178333759, + 0.0049293022602796555, + 0.004666678141802549, + 0.004119032062590122, + 0.006976825650781393, + 0.010807808488607407, + 0.004599603824317455, + 0.004159674979746342, + 0.007341632153838873, + 0.009681468829512596, + 0.007224144879728556, + 0.013471956364810467, + 0.008680024184286594, + 0.009808375500142574, + 0.008108768612146378, + 0.007163236383348703, + 0.014506304636597633, + 0.012332145124673843, + 0.015562302432954311, + 0.015486722812056541, + 0.022255701944231987, + 0.03477701544761658, + 0.014097761362791061, + 0.009667590260505676, + 0.03297540545463562, + 0.02145242504775524, + 0.008282379247248173, + 0.00855282787233591, + 0.006956785451620817, + 0.009083014912903309, + 0.0039609516970813274, + 0.0038365100044757128, + 0.016001034528017044, + 0.007316950708627701, + 0.007298520766198635, + 0.00840733852237463, + 0.00760316476225853, + 0.008731687441468239, + 0.004460249096155167, + 0.004777366295456886, + 0.009044786915183067, + 0.007725773844867945, + 0.01940971426665783, + 0.017789648845791817, + 0.023360321298241615, + 0.016564764082431793, + 0.009746639989316463, + 0.008998551405966282, + 0.017667395994067192, + 0.0180280152708292, + 0.006825852207839489, + 0.0076309917494654655, + 0.008544951677322388, + 0.009729682467877865, + 0.00609242357313633, + 0.005636649206280708, + 0.009452427737414837, + 0.0053132339380681515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.7720062732696533 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47436219453811646 + ], + "opposite_logprob_tokens": [ + -0.9743621945381165 + ], + "logprob_reward": -0.47436219453811646, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47436219453811646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1831977814435959, + 0.11382003128528595, + 0.07437995076179504, + 0.15367107093334198, + 0.08968597650527954, + 0.14661921560764313, + 0.04785316064953804, + 0.10626038908958435, + 0.04036084935069084, + 0.04415157064795494 + ], + "credit_map": [ + 0.024329151958227158, + 0.026043718680739403, + 0.03979874029755592, + 0.029542304575443268, + 0.013199017383158207, + 0.014489862136542797, + 0.01586473546922207, + 0.01993023045361042, + 0.016088198870420456, + 0.01800139993429184, + 0.02044646628201008, + 0.016437126323580742, + 0.009623569436371326, + 0.00920031126588583, + 0.01320524699985981, + 0.01081771682947874, + 0.00870959460735321, + 0.010820205323398113, + 0.010985173285007477, + 0.012251459993422031, + 0.0056776078417897224, + 0.006386011838912964, + 0.007514084689319134, + 0.012035808525979519, + 0.02577962353825569, + 0.011942483484745026, + 0.02344205230474472, + 0.03806615620851517, + 0.008975991979241371, + 0.007384893950074911, + 0.01941481977701187, + 0.018665049225091934, + 0.008748090825974941, + 0.01448440458625555, + 0.011298081837594509, + 0.015034941956400871, + 0.008716153912246227, + 0.006754137575626373, + 0.014822431840002537, + 0.009827733039855957, + 0.014733600430190563, + 0.014325900003314018, + 0.02024649642407894, + 0.035905759781599045, + 0.007456463761627674, + 0.007953081279993057, + 0.028528451919555664, + 0.01746946945786476, + 0.0036463632714003325, + 0.006640158127993345, + 0.0062194750644266605, + 0.01074560359120369, + 0.002589371521025896, + 0.0029637364204972982, + 0.009475710801780224, + 0.005572746507823467, + 0.014825353398919106, + 0.014351876452565193, + 0.013498667627573013, + 0.015677455812692642, + 0.007909147068858147, + 0.007412455044686794, + 0.016298972070217133, + 0.0162864588201046, + 0.0037902765907347202, + 0.004841235000640154, + 0.006924043409526348, + 0.006601027213037014, + 0.0029785253573209047, + 0.00345622724853456, + 0.006350588519126177, + 0.005418927408754826, + 0.0062187910079956055, + 0.005265464074909687, + 0.006427905056625605, + 0.007317105308175087, + 0.003433500649407506, + 0.0038429712876677513, + 0.007617082446813583, + 0.00402875104919076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.47436219453811646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036259638145565987 + ], + "opposite_logprob_tokens": [ + -5.628625869750977 + ], + "logprob_reward": -0.0036259638145565987, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036259638145565987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3735666871070862, + 0.14026854932308197, + 0.05208364874124527, + 0.06256704032421112, + 0.08279749751091003, + 0.09023697674274445, + 0.043515872210264206, + 0.043479274958372116, + 0.05970695987343788, + 0.05177750065922737 + ], + "credit_map": [ + 0.06210964918136597, + 0.0329788513481617, + 0.09511883556842804, + 0.04962827265262604, + 0.024750756099820137, + 0.02939840778708458, + 0.026377419009804726, + 0.053204495459795, + 0.014822755940258503, + 0.015179417096078396, + 0.028255056589841843, + 0.019587403163313866, + 0.015376639552414417, + 0.014947810210287571, + 0.016612427309155464, + 0.015487038530409336, + 0.0057731978595256805, + 0.00738378195092082, + 0.006324658170342445, + 0.00810618419200182, + 0.00613179337233305, + 0.0052825286984443665, + 0.005422717891633511, + 0.007658782880753279, + 0.005539746023714542, + 0.008049553260207176, + 0.006576916668564081, + 0.008236839435994625, + 0.005996727384626865, + 0.006917400285601616, + 0.008852431550621986, + 0.012397428043186665, + 0.010092203505337238, + 0.011220675893127918, + 0.011757990345358849, + 0.011907580308616161, + 0.004539629444479942, + 0.0063068135641515255, + 0.00959398690611124, + 0.017378611490130424, + 0.010904746130108833, + 0.009872853755950928, + 0.012998676858842373, + 0.01535023096948862, + 0.006781092379242182, + 0.0064137470908463, + 0.01695658080279827, + 0.01095905713737011, + 0.0044744014739990234, + 0.004848983138799667, + 0.005192119162529707, + 0.0069962432608008385, + 0.0027482258155941963, + 0.002956034615635872, + 0.011125473305583, + 0.005174390505999327, + 0.005065352190285921, + 0.006348080467432737, + 0.006747404579073191, + 0.0061638071201741695, + 0.0038058788049966097, + 0.004081076476722956, + 0.005602610297501087, + 0.005665065720677376, + 0.007508116774260998, + 0.008177205920219421, + 0.00960962101817131, + 0.008145506493747234, + 0.005350918974727392, + 0.004376372788101435, + 0.008463416248559952, + 0.008075802586972713, + 0.005017278250306845, + 0.008217701688408852, + 0.006586314179003239, + 0.007720206398516893, + 0.00517245102673769, + 0.004187291953712702, + 0.009980442002415657, + 0.004895814228802919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0036259638145565987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539473362267017 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539473362267017, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539473362267017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14157886803150177, + 0.17442160844802856, + 0.07542696595191956, + 0.06761989742517471, + 0.19009311497211456, + 0.10759608447551727, + 0.039798472076654434, + 0.08288206905126572, + 0.03889627009630203, + 0.08168665319681168 + ], + "credit_map": [ + 0.019588397815823555, + 0.014909299090504646, + 0.034479159861803055, + 0.020027490332722664, + 0.010947496630251408, + 0.010923249647021294, + 0.01788199692964554, + 0.012821780517697334, + 0.0242290198802948, + 0.009716039523482323, + 0.060907572507858276, + 0.03749336674809456, + 0.007768913637846708, + 0.008255822584033012, + 0.014165452681481838, + 0.011885430663824081, + 0.01335567981004715, + 0.0054633175022900105, + 0.012387530878186226, + 0.016020191833376884, + 0.00488481717184186, + 0.0052222865633666515, + 0.0082118920981884, + 0.009881251491606236, + 0.009545871987938881, + 0.006853011902421713, + 0.012991730123758316, + 0.013172194361686707, + 0.005188709124922752, + 0.004202245268970728, + 0.008328099735081196, + 0.007338032126426697, + 0.023991215974092484, + 0.013041768223047256, + 0.029824551194906235, + 0.030104998499155045, + 0.007965865544974804, + 0.017801905050873756, + 0.0169695857912302, + 0.0503932386636734, + 0.01568981446325779, + 0.009289072826504707, + 0.013965349644422531, + 0.017075933516025543, + 0.006836208514869213, + 0.006209561601281166, + 0.023215606808662415, + 0.015314539894461632, + 0.003944941330701113, + 0.003550716210156679, + 0.0049176206812262535, + 0.0072774761356413364, + 0.0022835610434412956, + 0.0030726962722837925, + 0.00911629106849432, + 0.005635166075080633, + 0.012477478943765163, + 0.014172178693115711, + 0.010208138264715672, + 0.01177535206079483, + 0.005770853254944086, + 0.006080503109842539, + 0.011321288533508778, + 0.011076281778514385, + 0.004405186511576176, + 0.004851116798818111, + 0.005963471252471209, + 0.006518373731523752, + 0.003143002511933446, + 0.00414033280685544, + 0.00652294373139739, + 0.0033518446143716574, + 0.012351475656032562, + 0.01237861905246973, + 0.011487581767141819, + 0.01382988877594471, + 0.006240243092179298, + 0.006276808213442564, + 0.008853673934936523, + 0.01026836410164833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.012539473362267017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20155513286590576 + ], + "opposite_logprob_tokens": [ + -1.7015551328659058 + ], + "logprob_reward": -0.20155513286590576, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155513286590576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16185148060321808, + 0.10747011005878448, + 0.08552718162536621, + 0.13073496520519257, + 0.11394636332988739, + 0.11083095520734787, + 0.12825459241867065, + 0.06003208085894585, + 0.04854707419872284, + 0.05280523747205734 + ], + "credit_map": [ + 0.026785986497998238, + 0.013778758235275745, + 0.047256968915462494, + 0.0226641446352005, + 0.011593721807003021, + 0.011732196435332298, + 0.014600410126149654, + 0.013439281843602657, + 0.013716344721615314, + 0.010195812210440636, + 0.02142041176557541, + 0.014103603549301624, + 0.010248739272356033, + 0.008992637507617474, + 0.016852395609021187, + 0.01194015797227621, + 0.012651047669351101, + 0.00838395580649376, + 0.016244469210505486, + 0.016383768990635872, + 0.007069359999150038, + 0.006177374627441168, + 0.008358110673725605, + 0.010259081609547138, + 0.012183433398604393, + 0.017489012330770493, + 0.01717211864888668, + 0.020623167976737022, + 0.013842690736055374, + 0.012696851044893265, + 0.015319655649363995, + 0.02140803076326847, + 0.014002725481987, + 0.011622507125139236, + 0.014407494105398655, + 0.02288593351840973, + 0.007994509302079678, + 0.008233831264078617, + 0.023103797808289528, + 0.011695549823343754, + 0.009711088612675667, + 0.01191038079559803, + 0.012984693050384521, + 0.013372752815485, + 0.010942365042865276, + 0.007662884425371885, + 0.02489178068935871, + 0.01935499720275402, + 0.011185020208358765, + 0.011175934225320816, + 0.014708012342453003, + 0.02112738788127899, + 0.008062111213803291, + 0.00927847158163786, + 0.0387696735560894, + 0.013947976753115654, + 0.006277511827647686, + 0.00601096311584115, + 0.007110827602446079, + 0.010707717388868332, + 0.0034638612996786833, + 0.004545473959296942, + 0.011426780372858047, + 0.01048894040286541, + 0.004703809507191181, + 0.007667002268135548, + 0.006611244287341833, + 0.006764674559235573, + 0.004925103858113289, + 0.0052145193330943584, + 0.00764269707724452, + 0.00501802284270525, + 0.004574631806463003, + 0.005828116554766893, + 0.008002882823348045, + 0.009123790077865124, + 0.005408843047916889, + 0.003780999220907688, + 0.009726145304739475, + 0.006359820254147053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.20155513286590576 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide-angle lens captures the scene from above the seafloor, the camera drifting to the right in unusually slow and steady lateral movements, creating a sense of reverent observation. In the profound azure silence of the deep ocean, a massive great white shark rests, hovering almost motionless above a pale, fine-sand seafloor dotted with gentle ripples from the current. Its massive body, a masterpiece of predatory evolution, is rendered in surreal detail; the slate-gray skin of its dorsal side, delicately marked with faint silver scars, contrasts starkly with the stark, clean white of its underbelly. Its black, hollow eyes gaze forward into the abyss, while its jaw, slightly ajar, offers a glimpse of jagged, triangular teeth. The only perceptible movement is the slow, rhythmic opening of its gill slits with breathing and the gentle, almost lazy drift of its powerful tail fin on the invisible current. Bright, ethereal sunlight penetrates the water from the distant surface, creating unique \"god rays\" that pierce the deep blue and illuminate swirling, dust-like particles of marine snow. These rays punctuate the shark’s powerful back, casting an ever-changing mosaic of light and shadow, highlighting its muscular silhouette and creating an atmosphere of tranquility yet tension.", + "seed": 26425, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_14/index_1461_seed_26425_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.188281749561429, + "vlm_logprob": -0.188281749561429, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.188281749561429, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5352654457092285 + ], + "opposite_logprob_tokens": [ + -0.9102654457092285 + ], + "logprob_reward": -0.5352654457092285, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5352654457092285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19177541136741638, + 0.16852223873138428, + 0.10033271461725235, + 0.1069437637925148, + 0.04515935853123665, + 0.060272231698036194, + 0.04904600977897644, + 0.1392052322626114, + 0.048210110515356064, + 0.09053295105695724 + ], + "credit_map": [ + 0.029485538601875305, + 0.016030356287956238, + 0.021557306870818138, + 0.023249264806509018, + 0.017032410949468613, + 0.015790294855833054, + 0.021488651633262634, + 0.04714159294962883, + 0.01744341105222702, + 0.01189547497779131, + 0.015983421355485916, + 0.024535031989216805, + 0.012695899233222008, + 0.01149191614240408, + 0.019433245062828064, + 0.05504383519291878, + 0.01084135565906763, + 0.008540941402316093, + 0.012665840797126293, + 0.013032572343945503, + 0.006837681867182255, + 0.008044891059398651, + 0.011769270524382591, + 0.028600160032510757, + 0.01074003241956234, + 0.012211662717163563, + 0.01533167902380228, + 0.014593277126550674, + 0.011958618648350239, + 0.017366930842399597, + 0.010379711166024208, + 0.0143618518486619, + 0.005470681469887495, + 0.007091642823070288, + 0.005576727911829948, + 0.006987405009567738, + 0.005730705335736275, + 0.004854839760810137, + 0.004038759041577578, + 0.005408598110079765, + 0.007291561458259821, + 0.007128796074539423, + 0.00878528505563736, + 0.010029769502580166, + 0.005910290405154228, + 0.005679647903889418, + 0.006423281505703926, + 0.009023597463965416, + 0.004615457262843847, + 0.00822929386049509, + 0.00699474336579442, + 0.00862858910113573, + 0.005527789238840342, + 0.006115348543971777, + 0.003965223673731089, + 0.00496956380084157, + 0.01849386654794216, + 0.011782870627939701, + 0.027608001604676247, + 0.04009956121444702, + 0.010018525645136833, + 0.009501447901129723, + 0.01390835177153349, + 0.007792619522660971, + 0.0065584671683609486, + 0.006369674112647772, + 0.006563121918588877, + 0.008175999857485294, + 0.005592869129031897, + 0.0057215094566345215, + 0.00475804740563035, + 0.00447042053565383, + 0.014442124404013157, + 0.0066633401438593864, + 0.020498789846897125, + 0.026038603857159615, + 0.003998436499387026, + 0.00455063208937645, + 0.0075795818120241165, + 0.0067614419385790825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.5352654457092285 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159585773944855 + ], + "opposite_logprob_tokens": [ + -1.7015959024429321 + ], + "logprob_reward": -0.20159585773944855, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159585773944855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06944401562213898, + 0.18575547635555267, + 0.13958144187927246, + 0.06504248827695847, + 0.04455985128879547, + 0.07606428116559982, + 0.12644237279891968, + 0.1553996056318283, + 0.025084946304559708, + 0.11262550204992294 + ], + "credit_map": [ + 0.011408085003495216, + 0.00724790757521987, + 0.0072662681341171265, + 0.009620442055165768, + 0.006789523176848888, + 0.007091302890330553, + 0.009992502629756927, + 0.010027988813817501, + 0.020507220178842545, + 0.011217215098440647, + 0.01869770884513855, + 0.023516783490777016, + 0.00885179452598095, + 0.01131705567240715, + 0.022373497486114502, + 0.06927421689033508, + 0.017131812870502472, + 0.011032110080122948, + 0.02033301815390587, + 0.02116725780069828, + 0.007049183361232281, + 0.009966140612959862, + 0.01828625053167343, + 0.03461567685008049, + 0.008363462053239346, + 0.0062933024019002914, + 0.008962828665971756, + 0.012977235019207, + 0.005285824183374643, + 0.005122706294059753, + 0.007091217674314976, + 0.010945913381874561, + 0.007896187715232372, + 0.004594160709530115, + 0.007070858497172594, + 0.007201496511697769, + 0.003426001640036702, + 0.0031551853753626347, + 0.004832408856600523, + 0.006383554544299841, + 0.011449780315160751, + 0.007396813947707415, + 0.012755492702126503, + 0.015823353081941605, + 0.00588825810700655, + 0.005850475747138262, + 0.00706627918407321, + 0.009833830408751965, + 0.01342584379017353, + 0.021238014101982117, + 0.017565803602337837, + 0.024958567693829536, + 0.013885789550840855, + 0.013677502982318401, + 0.01095503382384777, + 0.010735822841525078, + 0.019549330696463585, + 0.012478700838983059, + 0.03176990523934364, + 0.045936889946460724, + 0.011074641719460487, + 0.00978038925677538, + 0.015890777111053467, + 0.00891898013651371, + 0.003678517648950219, + 0.0017926553264260292, + 0.00624260725453496, + 0.005525726359337568, + 0.001808410743251443, + 0.0017580537823960185, + 0.002197406254708767, + 0.0020815699826925993, + 0.018078280612826347, + 0.009049559012055397, + 0.026443278416991234, + 0.03155522793531418, + 0.005302985664457083, + 0.005601809825748205, + 0.008709593676030636, + 0.00788477249443531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.20159585773944855 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101555794477463 + ], + "opposite_logprob_tokens": [ + -5.504101753234863 + ], + "logprob_reward": -0.004101555794477463, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101555794477463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42752137780189514, + 0.09775698184967041, + 0.09158823639154434, + 0.06237073987722397, + 0.07837595790624619, + 0.04201893508434296, + 0.06793728470802307, + 0.04871758446097374, + 0.03068746067583561, + 0.053025417029857635 + ], + "credit_map": [ + 0.041321463882923126, + 0.03313819319009781, + 0.04853108152747154, + 0.053465038537979126, + 0.03634887933731079, + 0.03569638729095459, + 0.04709182679653168, + 0.13192850351333618, + 0.009244619868695736, + 0.011054256930947304, + 0.010460286401212215, + 0.013473658822476864, + 0.010468660853803158, + 0.009525771252810955, + 0.013601343147456646, + 0.019928397610783577, + 0.008724614977836609, + 0.010240391828119755, + 0.01077299751341343, + 0.01203078031539917, + 0.009016748517751694, + 0.006680299062281847, + 0.010986329056322575, + 0.02313607558608055, + 0.006602735724300146, + 0.00914890319108963, + 0.007688689511269331, + 0.009798354469239712, + 0.006439198739826679, + 0.006278070621192455, + 0.006493217311799526, + 0.009921577759087086, + 0.008550903759896755, + 0.01504090242087841, + 0.009350805543363094, + 0.00790884904563427, + 0.009882625192403793, + 0.00739577179774642, + 0.007448130287230015, + 0.012797976844012737, + 0.003949882462620735, + 0.004432788118720055, + 0.005082447547465563, + 0.005240504164248705, + 0.006583344656974077, + 0.007102569565176964, + 0.0037531426642090082, + 0.005874260328710079, + 0.005990007892251015, + 0.007866941392421722, + 0.00988336093723774, + 0.01231314055621624, + 0.009023802354931831, + 0.007389487233012915, + 0.0077182212844491005, + 0.007752329111099243, + 0.004548126365989447, + 0.0036548322532325983, + 0.008514025248587132, + 0.012463021092116833, + 0.00558749632909894, + 0.004315892234444618, + 0.0063563743606209755, + 0.003277821699157357, + 0.00448707677423954, + 0.004241796676069498, + 0.0045464420691132545, + 0.005335177294909954, + 0.0037022223696112633, + 0.003192496718838811, + 0.0024766302667558193, + 0.0027056203689426184, + 0.0050681703723967075, + 0.006063163746148348, + 0.006820730399340391, + 0.007960519753396511, + 0.006777164060622454, + 0.005357846152037382, + 0.007017047144472599, + 0.007960784249007702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004101555794477463 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579038187861443 + ], + "opposite_logprob_tokens": [ + -3.8955790996551514 + ], + "logprob_reward": -0.020579038187861443, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579038187861443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19652780890464783, + 0.13947001099586487, + 0.11089997738599777, + 0.07523882389068604, + 0.03421003744006157, + 0.06616892665624619, + 0.0957699716091156, + 0.21085916459560394, + 0.02017003484070301, + 0.05068519711494446 + ], + "credit_map": [ + 0.021571725606918335, + 0.01774907484650612, + 0.024465374648571014, + 0.02829723432660103, + 0.01786169223487377, + 0.016180461272597313, + 0.023409321904182434, + 0.04699292033910751, + 0.01432896964251995, + 0.009886150248348713, + 0.015170399099588394, + 0.020368024706840515, + 0.009498952887952328, + 0.010599701665341854, + 0.017262833192944527, + 0.04235498607158661, + 0.010561899282038212, + 0.006653955206274986, + 0.012873804196715355, + 0.018489226698875427, + 0.005071548279374838, + 0.008803254924714565, + 0.014158002100884914, + 0.03428828343749046, + 0.010655606165528297, + 0.006254235748201609, + 0.011173458769917488, + 0.017568834125995636, + 0.00501558231189847, + 0.006253700237721205, + 0.006893866695463657, + 0.011423536576330662, + 0.00483986409381032, + 0.004162555560469627, + 0.004096247721463442, + 0.005960062146186829, + 0.0037368631456047297, + 0.0038978541269898415, + 0.0035208063200116158, + 0.003995785489678383, + 0.008858656510710716, + 0.005369915161281824, + 0.012177526019513607, + 0.018587199971079826, + 0.0055642216466367245, + 0.0053277816623449326, + 0.0048391493037343025, + 0.005444479174911976, + 0.014896295964717865, + 0.009949601255357265, + 0.018606631085276604, + 0.02345960959792137, + 0.007672583218663931, + 0.005043307319283485, + 0.007502786349505186, + 0.008639167994260788, + 0.025421304628252983, + 0.018603598698973656, + 0.045800358057022095, + 0.05702923983335495, + 0.014889631420373917, + 0.012299581430852413, + 0.023353759199380875, + 0.013461701571941376, + 0.003199484897777438, + 0.002082295948639512, + 0.0034495415166020393, + 0.004044997505843639, + 0.0020827993284910917, + 0.0018081640591844916, + 0.0022129511926323175, + 0.0012897998094558716, + 0.0069071161560714245, + 0.005208706017583609, + 0.00780053436756134, + 0.009507471695542336, + 0.005362547468394041, + 0.004770916420966387, + 0.005784165114164352, + 0.005343737546354532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.020579038187861443 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle looking down from above the seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive great white shark visible hovering above a pale, fine-sand seafloor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the shark have slate-gray skin on its dorsal side with faint silver scars and a stark white underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright ethereal sunlight rays (god rays) visible penetrating the water and illuminating marine snow particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera drift slowly and steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986685037612915 + ], + "opposite_logprob_tokens": [ + -1.8048667907714844 + ], + "logprob_reward": -0.17986685037612915, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986685037612915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11425565183162689, + 0.13277111947536469, + 0.09893914312124252, + 0.09049132466316223, + 0.034955885261297226, + 0.11470206826925278, + 0.13874328136444092, + 0.07101522386074066, + 0.048047833144664764, + 0.15607850253582 + ], + "credit_map": [ + 0.016416048631072044, + 0.010691896080970764, + 0.019356006756424904, + 0.015493051148951054, + 0.009714134968817234, + 0.009094011038541794, + 0.012511903420090675, + 0.020978599786758423, + 0.01610701158642769, + 0.01020343042910099, + 0.013880149461328983, + 0.016882043331861496, + 0.009257156401872635, + 0.008376166224479675, + 0.015157436020672321, + 0.0429077111184597, + 0.009949568659067154, + 0.011543265543878078, + 0.010572202503681183, + 0.012232922948896885, + 0.00914851576089859, + 0.01023456733673811, + 0.015454615466296673, + 0.019803486764431, + 0.009614720940589905, + 0.007728241384029388, + 0.011190960183739662, + 0.014321940951049328, + 0.006946186069399118, + 0.00868212804198265, + 0.011561861261725426, + 0.02044527977705002, + 0.0031662036199122667, + 0.00514242984354496, + 0.004146430175751448, + 0.0057479641400277615, + 0.004213706590235233, + 0.0028833947144448757, + 0.004028582014143467, + 0.00562717579305172, + 0.015445169061422348, + 0.010396953672170639, + 0.017956072464585304, + 0.019786464050412178, + 0.00853612832725048, + 0.008940943516790867, + 0.012074504978954792, + 0.021565834060311317, + 0.013950907625257969, + 0.025789299979805946, + 0.018910778686404228, + 0.023943182080984116, + 0.015509510412812233, + 0.018787473440170288, + 0.011234409175813198, + 0.010617717169225216, + 0.006848767399787903, + 0.006030784919857979, + 0.012569023296236992, + 0.017179636284708977, + 0.00724708940833807, + 0.009877894073724747, + 0.006092211697250605, + 0.005169819574803114, + 0.006698442157357931, + 0.004574083257466555, + 0.006537822540849447, + 0.00868888758122921, + 0.006574430502951145, + 0.006538955960422754, + 0.004601763561367989, + 0.0038334501441568136, + 0.0163543913513422, + 0.02074536494910717, + 0.022971423342823982, + 0.025221489369869232, + 0.019109882414340973, + 0.018051737919449806, + 0.012990070506930351, + 0.020634153857827187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.17986685037612915 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_15/index_tensor([3713, 3713])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_15/index_tensor([3713, 3713])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7980a2d319689c1117f896583ab4086b15f0b4f7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_15/index_tensor([3713, 3713])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium side shot, rendered as a hyper-realistic CGI animation, tracks smoothly alongside the massive Atlantean city-ship as it descends majestically into the conformist black void of the ocean abyss. The camera maintains a steady, parallel path, tilting slowly downward to emphasize the city's immense scale and its steady, inexorable downward momentum. The city's architecture is a breathtaking fusion of organic forms and advanced technology. Towering crystalline spires shimmer with a soft internal bioluminescence, casting an ethereal divine light of turquoise and gold into the surrounding darkness. Waterfalls appear to originate from within the city, cascading silently down steps, streams catching the inner light and shimmering like liquid diamonds. In stunning, physically accurate detail, the dark waters surround the city's massive, streamlined hull, creating a visible shell of compressed, supercavitated space, seething with microscopic bubbles, while the flow along its flanks appears perfectly laminar and smooth. The scene is filled with an atmosphere of silent awe, capturing a moment of impossible grandeur and ancient mystery with lifelike precision, as clouds of illuminated marine snow drift past like a slow-motion star field.", + "seed": 20495, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_15/index_3713_seed_20495_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02249199866782874, + "vlm_logprob": -0.02249199866782874, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.02249199866782874, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002380040008574724 + ], + "opposite_logprob_tokens": [ + -6.627379894256592 + ], + "logprob_reward": -0.002380040008574724, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.002380040008574724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13446861505508423, + 0.1588456779718399, + 0.08173732459545135, + 0.0737738236784935, + 0.31093305349349976, + 0.06455269455909729, + 0.025977808982133865, + 0.05623411387205124, + 0.04065030813217163, + 0.052826572209596634 + ], + "credit_map": [ + 0.01551967952400446, + 0.020066631957888603, + 0.015060980804264545, + 0.009513693861663342, + 0.01846097595989704, + 0.016144173219799995, + 0.024591587483882904, + 0.015110903419554234, + 0.02349013276398182, + 0.026931149885058403, + 0.01791987381875515, + 0.012785397469997406, + 0.023063702508807182, + 0.013444402255117893, + 0.018514445051550865, + 0.022696558386087418, + 0.01510502677410841, + 0.007496999576687813, + 0.01044709887355566, + 0.006410487927496433, + 0.007458746898919344, + 0.011472349055111408, + 0.012906679883599281, + 0.01043994352221489, + 0.013122506439685822, + 0.013653420843183994, + 0.008453796617686749, + 0.005584020633250475, + 0.007325328886508942, + 0.00678078830242157, + 0.008162928745150566, + 0.010691026225686073, + 0.04594758525490761, + 0.05959971621632576, + 0.07584959268569946, + 0.017883123829960823, + 0.04974568635225296, + 0.018920324742794037, + 0.01676386408507824, + 0.02622315101325512, + 0.006725771818310022, + 0.0094752237200737, + 0.010768565349280834, + 0.009784761816263199, + 0.005680173635482788, + 0.008945271372795105, + 0.004321385640650988, + 0.008851545862853527, + 0.00332690984942019, + 0.0028370895888656378, + 0.0037183244712650776, + 0.0025312043726444244, + 0.004112311638891697, + 0.003186777699738741, + 0.0029799097683280706, + 0.0032852841541171074, + 0.004672930110245943, + 0.006808838807046413, + 0.005515492986887693, + 0.004405893851071596, + 0.006346223410218954, + 0.008568599820137024, + 0.01187082938849926, + 0.008045303635299206, + 0.008212710730731487, + 0.004702605772763491, + 0.005718591157346964, + 0.00541111221536994, + 0.004134884104132652, + 0.004082949366420507, + 0.004233676940202713, + 0.00415377551689744, + 0.004819221794605255, + 0.007070187479257584, + 0.0037363520823419094, + 0.007380644790828228, + 0.006999643053859472, + 0.007644353434443474, + 0.006798328831791878, + 0.008377843536436558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.002380040008574724 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013446108205243945 + ], + "opposite_logprob_tokens": [ + -6.626344680786133 + ], + "logprob_reward": -0.0013446108205243945, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013446108205243945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0704977884888649, + 0.051309797912836075, + 0.04254557192325592, + 0.07838569581508636, + 0.41495680809020996, + 0.05267532542347908, + 0.03653615713119507, + 0.11361117660999298, + 0.0883932113647461, + 0.05108850076794624 + ], + "credit_map": [ + 0.008060180582106113, + 0.010356719605624676, + 0.010258307680487633, + 0.006728726904839277, + 0.00867205299437046, + 0.008638947270810604, + 0.010619751177728176, + 0.007163103204220533, + 0.005174883641302586, + 0.007709338795393705, + 0.0049566966481506824, + 0.007362773641943932, + 0.00781684834510088, + 0.005153432488441467, + 0.00729431863874197, + 0.005841503385454416, + 0.0067727589048445225, + 0.0052583059296011925, + 0.0051580569706857204, + 0.003626827849075198, + 0.004964616149663925, + 0.005214208271354437, + 0.006903775501996279, + 0.004647024907171726, + 0.014741797000169754, + 0.014841779135167599, + 0.007042760029435158, + 0.005660733673721552, + 0.011863497085869312, + 0.006047772243618965, + 0.00673518655821681, + 0.011452173814177513, + 0.06579338014125824, + 0.06773793697357178, + 0.1249275729060173, + 0.022842317819595337, + 0.05414861813187599, + 0.02174603007733822, + 0.022022509947419167, + 0.03573844954371452, + 0.005634868983179331, + 0.006834035739302635, + 0.009111122228205204, + 0.005959307309240103, + 0.004201636184006929, + 0.0069439103826880455, + 0.00443480908870697, + 0.009555638767778873, + 0.0022983807139098644, + 0.0034613788593560457, + 0.009939643554389477, + 0.003121278015896678, + 0.003657909110188484, + 0.0025790215004235506, + 0.0031533571891486645, + 0.008325187489390373, + 0.0122278593480587, + 0.01375490427017212, + 0.011828605085611343, + 0.011234823614358902, + 0.011660385876893997, + 0.011387658305466175, + 0.0241550300270319, + 0.017361899837851524, + 0.013412301428616047, + 0.009112498722970486, + 0.01225240621715784, + 0.013003823347389698, + 0.00806584395468235, + 0.011039238423109055, + 0.011255187913775444, + 0.010251904837787151, + 0.0048260195180773735, + 0.007498724851757288, + 0.004026839975267649, + 0.0061327675357460976, + 0.008848389610648155, + 0.0068897223100066185, + 0.005555704701691866, + 0.007310332730412483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.0013446108205243945 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636622689664364 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636622689664364, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636622689664364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12947535514831543, + 0.05916719511151314, + 0.043733447790145874, + 0.06434480100870132, + 0.32363566756248474, + 0.03277044743299484, + 0.05908671021461487, + 0.066194087266922, + 0.13535501062870026, + 0.08623729646205902 + ], + "credit_map": [ + 0.012918655760586262, + 0.017956459894776344, + 0.015598569065332413, + 0.012142553925514221, + 0.015281527303159237, + 0.012491590343415737, + 0.028799334540963173, + 0.014286667108535767, + 0.0072082458063960075, + 0.00850165169686079, + 0.007029276806861162, + 0.007005335763096809, + 0.005155705846846104, + 0.00546726631000638, + 0.008956076577305794, + 0.009843635372817516, + 0.0070358882658183575, + 0.005043196491897106, + 0.005234569311141968, + 0.0037102929782122374, + 0.004806975834071636, + 0.004286999814212322, + 0.006559485103935003, + 0.007056039292365313, + 0.01009383425116539, + 0.009819835424423218, + 0.008029457181692123, + 0.007037153467535973, + 0.005852686706930399, + 0.00673781568184495, + 0.008710308000445366, + 0.00806371122598648, + 0.04913118854165077, + 0.05779823660850525, + 0.08802960067987442, + 0.0182966198772192, + 0.04748758673667908, + 0.017509521916508675, + 0.01749325543642044, + 0.027889661490917206, + 0.004051131196320057, + 0.005403608549386263, + 0.004435576498508453, + 0.004638367332518101, + 0.003164761932566762, + 0.003727831644937396, + 0.002765633864328265, + 0.004583535250276327, + 0.004811879713088274, + 0.00463114446029067, + 0.015592747367918491, + 0.005399672314524651, + 0.005456744693219662, + 0.00704159215092659, + 0.005544416606426239, + 0.010608515702188015, + 0.005648504942655563, + 0.006902588065713644, + 0.009276806376874447, + 0.006891102530062199, + 0.005382386036217213, + 0.006101319100707769, + 0.016406545415520668, + 0.009584831073880196, + 0.01960359886288643, + 0.015362484380602837, + 0.017323695123195648, + 0.012774421833455563, + 0.015852436423301697, + 0.02297016605734825, + 0.01540639903396368, + 0.0160618107765913, + 0.009508366696536541, + 0.011179736815392971, + 0.01120592188090086, + 0.010203342884778976, + 0.010376328602433205, + 0.009941257536411285, + 0.012270965613424778, + 0.011551383882761002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.008636622689664364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902539312839508 + ], + "opposite_logprob_tokens": [ + -2.4640254974365234 + ], + "logprob_reward": -0.08902539312839508, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902539312839508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09807083755731583, + 0.10213848203420639, + 0.05512632057070732, + 0.1408771127462387, + 0.24030297994613647, + 0.026759978383779526, + 0.05872967466711998, + 0.070996955037117, + 0.10362017899751663, + 0.10337755084037781 + ], + "credit_map": [ + 0.009654093533754349, + 0.014426966197788715, + 0.012514747679233551, + 0.00928754173219204, + 0.01217611413449049, + 0.011926956474781036, + 0.01700625754892826, + 0.011078163050115108, + 0.01233693864196539, + 0.014870346523821354, + 0.015681715682148933, + 0.009393713437020779, + 0.015755221247673035, + 0.009395486675202847, + 0.012428668327629566, + 0.012276390567421913, + 0.008639540523290634, + 0.005796171259135008, + 0.0073709674179553986, + 0.005391561891883612, + 0.005726559087634087, + 0.008057664148509502, + 0.008209418505430222, + 0.005934440530836582, + 0.024998096749186516, + 0.03087172657251358, + 0.012705466710031033, + 0.012486791238188744, + 0.017193196341395378, + 0.009873435832560062, + 0.01656077802181244, + 0.01618761382997036, + 0.038495805114507675, + 0.03138696402311325, + 0.07592268288135529, + 0.013990613631904125, + 0.021386845037341118, + 0.01168328057974577, + 0.019134512171149254, + 0.02830228954553604, + 0.0027153068222105503, + 0.0026743679773062468, + 0.0069029973819851875, + 0.003575497306883335, + 0.001900894450955093, + 0.0018465403700247407, + 0.0021567128133028746, + 0.00498766265809536, + 0.0034041397739201784, + 0.003851112211123109, + 0.016243135556578636, + 0.006379709579050541, + 0.004087660927325487, + 0.004148344974964857, + 0.006894165650010109, + 0.013721403665840626, + 0.006559740751981735, + 0.007538113743066788, + 0.007886347360908985, + 0.006092667579650879, + 0.007846578024327755, + 0.008393687196075916, + 0.016096588224172592, + 0.010583232156932354, + 0.016948474571108818, + 0.008800028823316097, + 0.014276113361120224, + 0.016124865040183067, + 0.009647045284509659, + 0.011347153224050999, + 0.013380512595176697, + 0.01309597585350275, + 0.011278323829174042, + 0.011705689132213593, + 0.01499233115464449, + 0.012658408842980862, + 0.009706850163638592, + 0.010639171116054058, + 0.017787914723157883, + 0.01460886001586914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.08902539312839508 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107332669198513 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.01107332669198513, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107332669198513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10056496411561966, + 0.06763727217912674, + 0.047112803906202316, + 0.0847676545381546, + 0.37424665689468384, + 0.04225607216358185, + 0.04811583831906319, + 0.05447075515985489, + 0.09463731944561005, + 0.08619069308042526 + ], + "credit_map": [ + 0.009821404702961445, + 0.011982573196291924, + 0.009722852148115635, + 0.008915356360375881, + 0.014837154187262058, + 0.010048357769846916, + 0.02515440061688423, + 0.010082856751978397, + 0.010158858262002468, + 0.010693810880184174, + 0.00741765508428216, + 0.007387953344732523, + 0.009062975645065308, + 0.006144714076071978, + 0.009881864301860332, + 0.006889427546411753, + 0.007398710586130619, + 0.006189244333654642, + 0.006634997203946114, + 0.004583256784826517, + 0.007145846728235483, + 0.003915979526937008, + 0.0054473173804581165, + 0.005797447636723518, + 0.013141737319529057, + 0.013425406068563461, + 0.009537679143249989, + 0.00856305193156004, + 0.010560582391917706, + 0.007202954962849617, + 0.012110271491110325, + 0.010225964710116386, + 0.058076292276382446, + 0.05924658849835396, + 0.11567806452512741, + 0.020253939554095268, + 0.04794510453939438, + 0.020010458305478096, + 0.01993327960371971, + 0.03310288116335869, + 0.0045565892942249775, + 0.0065352050587534904, + 0.005972274113446474, + 0.00664868438616395, + 0.004467411432415247, + 0.005516160279512405, + 0.003048291662707925, + 0.005511448252946138, + 0.0034175741020590067, + 0.0042267837561666965, + 0.01207406260073185, + 0.004737993236631155, + 0.003893006592988968, + 0.0046114311553537846, + 0.005359656177461147, + 0.00979532115161419, + 0.0047028884291648865, + 0.005439355969429016, + 0.00886058621108532, + 0.005356790963560343, + 0.0051141162402927876, + 0.004929881542921066, + 0.012750542722642422, + 0.007316582836210728, + 0.01119360513985157, + 0.008744685910642147, + 0.012565622106194496, + 0.016251597553491592, + 0.011169057339429855, + 0.009058978408575058, + 0.013135844841599464, + 0.01251792348921299, + 0.007728149648755789, + 0.010540518909692764, + 0.010314463637769222, + 0.010975132696330547, + 0.008900781162083149, + 0.013284772634506226, + 0.011308226734399796, + 0.013138635084033012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.01107332669198513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium side shot, rendered as a hyper-realistic CGI animation, tracks smoothly alongside the massive Atlantean city-ship as it descends majestically into the conformist black void of the ocean abyss. The camera maintains a steady, parallel path, tilting slowly downward to emphasize the city's immense scale and its steady, inexorable downward momentum. The city's architecture is a breathtaking fusion of organic forms and advanced technology. Towering crystalline spires shimmer with a soft internal bioluminescence, casting an ethereal divine light of turquoise and gold into the surrounding darkness. Waterfalls appear to originate from within the city, cascading silently down steps, streams catching the inner light and shimmering like liquid diamonds. In stunning, physically accurate detail, the dark waters surround the city's massive, streamlined hull, creating a visible shell of compressed, supercavitated space, seething with microscopic bubbles, while the flow along its flanks appears perfectly laminar and smooth. The scene is filled with an atmosphere of silent awe, capturing a moment of impossible grandeur and ancient mystery with lifelike precision, as clouds of illuminated marine snow drift past like a slow-motion star field.", + "seed": 20496, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_15/index_3713_seed_20496_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06512770438566803, + "vlm_logprob": -0.06512770438566803, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.06512770438566803, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0661095604300499 + ], + "opposite_logprob_tokens": [ + -2.8161096572875977 + ], + "logprob_reward": -0.0661095604300499, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0661095604300499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19173546135425568, + 0.0657457634806633, + 0.13083206117153168, + 0.13574744760990143, + 0.033746685832738876, + 0.09805125743150711, + 0.11399902403354645, + 0.0973343700170517, + 0.05996682867407799, + 0.07284105569124222 + ], + "credit_map": [ + 0.01480024866759777, + 0.015516244806349277, + 0.020782470703125, + 0.024387694895267487, + 0.017809182405471802, + 0.029635600745677948, + 0.03270294517278671, + 0.036101073026657104, + 0.005212712567299604, + 0.009149468503892422, + 0.0065342728048563, + 0.007566694635897875, + 0.00899452157318592, + 0.007991662248969078, + 0.00952740479260683, + 0.010769030079245567, + 0.005128818564116955, + 0.005482516251504421, + 0.009480386041104794, + 0.013046487234532833, + 0.00627053901553154, + 0.013814233243465424, + 0.04144219681620598, + 0.03616689145565033, + 0.010668013244867325, + 0.014258695766329765, + 0.014328082092106342, + 0.013592296279966831, + 0.014937343075871468, + 0.02429647371172905, + 0.01794372871518135, + 0.025722818449139595, + 0.0029820504132658243, + 0.0027544014155864716, + 0.0026701735332608223, + 0.003348623402416706, + 0.005427227821201086, + 0.005353289656341076, + 0.00559005094692111, + 0.005620868410915136, + 0.009569646790623665, + 0.009935629554092884, + 0.009057585150003433, + 0.012248058803379536, + 0.016370870172977448, + 0.016689877957105637, + 0.011028533801436424, + 0.013151053339242935, + 0.012082771398127079, + 0.00912830512970686, + 0.00946737825870514, + 0.007448713295161724, + 0.021275920793414116, + 0.020749814808368683, + 0.016223961487412453, + 0.01762215979397297, + 0.014248586259782314, + 0.010011259466409683, + 0.00692426273599267, + 0.007135156076401472, + 0.02329416014254093, + 0.02327236346900463, + 0.005106483120471239, + 0.007342101540416479, + 0.00931878387928009, + 0.005689515266567469, + 0.004682071041315794, + 0.00499127060174942, + 0.013135413639247417, + 0.012428835965692997, + 0.004448520950973034, + 0.0052724191918969154, + 0.011715707369148731, + 0.00954530667513609, + 0.008369424380362034, + 0.009585489518940449, + 0.00856799352914095, + 0.009775487706065178, + 0.008554598316550255, + 0.006727045867592096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0661095604300499 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007650243584066629 + ], + "opposite_logprob_tokens": [ + -4.882650375366211 + ], + "logprob_reward": -0.007650243584066629, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007650243584066629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19910229742527008, + 0.10782862454652786, + 0.1838594675064087, + 0.22488516569137573, + 0.027873460203409195, + 0.03886115178465843, + 0.06459318101406097, + 0.048476286232471466, + 0.05863776430487633, + 0.045882634818553925 + ], + "credit_map": [ + 0.014889461919665337, + 0.021853210404515266, + 0.026986028999090195, + 0.019709058105945587, + 0.022098511457443237, + 0.024419235065579414, + 0.031985871493816376, + 0.03716091439127922, + 0.011079486459493637, + 0.011207097209990025, + 0.00890068244189024, + 0.009931112639605999, + 0.015794798731803894, + 0.01466350443661213, + 0.017699839547276497, + 0.018552109599113464, + 0.004130751825869083, + 0.004565461538732052, + 0.005494812503457069, + 0.01395337749272585, + 0.006415859796106815, + 0.00972652155905962, + 0.07143145054578781, + 0.06814121454954147, + 0.017697161063551903, + 0.03668751195073128, + 0.020770035684108734, + 0.020206501707434654, + 0.02546757645905018, + 0.0341062918305397, + 0.03864404931664467, + 0.03130602836608887, + 0.0024969466030597687, + 0.0020379892084747553, + 0.0021942988969385624, + 0.0028545435052365065, + 0.005299748387187719, + 0.004684137180447578, + 0.004176151938736439, + 0.004129643086344004, + 0.003595234826207161, + 0.0033282299991697073, + 0.0029611578211188316, + 0.00508245499804616, + 0.00830748025327921, + 0.007089464925229549, + 0.004658232908695936, + 0.0038388927932828665, + 0.0052936808206140995, + 0.004392421338707209, + 0.003569237422198057, + 0.003274857997894287, + 0.017849059775471687, + 0.016917968168854713, + 0.006224094424396753, + 0.007071861065924168, + 0.004807960242033005, + 0.008296971209347248, + 0.0028566867113113403, + 0.0041907792910933495, + 0.010221407748758793, + 0.010034792125225067, + 0.0035947435535490513, + 0.004472949076443911, + 0.00912580918520689, + 0.005541116930544376, + 0.004544912837445736, + 0.006743210833519697, + 0.0068096681497991085, + 0.007733291946351528, + 0.010053489357233047, + 0.008086263202130795, + 0.00786602683365345, + 0.003740954678505659, + 0.00592252379283309, + 0.004339747130870819, + 0.006036115810275078, + 0.007721246685832739, + 0.005941658280789852, + 0.004314362537115812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.007650243584066629 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048668745905160904 + ], + "opposite_logprob_tokens": [ + -3.04866886138916 + ], + "logprob_reward": -0.048668745905160904, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048668745905160904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06504151970148087, + 0.15073102712631226, + 0.14304572343826294, + 0.09553968906402588, + 0.05140791833400726, + 0.06956782191991806, + 0.13190661370754242, + 0.07195275276899338, + 0.1459522694349289, + 0.07485472410917282 + ], + "credit_map": [ + 0.007537355646491051, + 0.005684089381247759, + 0.006462602876126766, + 0.008673780597746372, + 0.007643985096365213, + 0.008849916979670525, + 0.010535775683820248, + 0.009654000401496887, + 0.013961679302155972, + 0.018520040437579155, + 0.012952934019267559, + 0.01528963539749384, + 0.019419947639107704, + 0.016962764784693718, + 0.025042571127414703, + 0.028581425547599792, + 0.00416564429178834, + 0.00523298978805542, + 0.005923515185713768, + 0.00977081898599863, + 0.007256234530359507, + 0.0117776645347476, + 0.05218353495001793, + 0.04673530161380768, + 0.0068682278506457806, + 0.0072104004211723804, + 0.00960309524089098, + 0.008600574918091297, + 0.008872080594301224, + 0.022533221170306206, + 0.010310274548828602, + 0.021541796624660492, + 0.004499319940805435, + 0.0033653187565505505, + 0.004530203528702259, + 0.0055013434030115604, + 0.00847192108631134, + 0.0083525525406003, + 0.00813357625156641, + 0.008553667925298214, + 0.007932992652058601, + 0.0076155830174684525, + 0.006551622413098812, + 0.007048388477414846, + 0.01205886248499155, + 0.01180973183363676, + 0.008430915884673595, + 0.008119719102978706, + 0.019667422398924828, + 0.012553568929433823, + 0.011426509357988834, + 0.008532986976206303, + 0.01950579136610031, + 0.01843879558146, + 0.019191449508070946, + 0.02259007655084133, + 0.006462387274950743, + 0.010271591134369373, + 0.005210238043218851, + 0.006563168950378895, + 0.012393761426210403, + 0.01275022141635418, + 0.008083757944405079, + 0.010217616334557533, + 0.021669624373316765, + 0.015145699493587017, + 0.013488433323800564, + 0.013845536857843399, + 0.014297588728368282, + 0.019750025123357773, + 0.02411426045000553, + 0.023641075938940048, + 0.01082354225218296, + 0.006662086118012667, + 0.010349069721996784, + 0.007009749300777912, + 0.009800994768738747, + 0.010753124952316284, + 0.011761944741010666, + 0.007694200612604618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.048668745905160904 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992201447486877 + ], + "opposite_logprob_tokens": [ + -1.8049219846725464 + ], + "logprob_reward": -0.17992201447486877, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992201447486877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04645382985472679, + 0.0423637330532074, + 0.13591094315052032, + 0.07448197901248932, + 0.029513001441955566, + 0.05120822414755821, + 0.13139373064041138, + 0.07440605014562607, + 0.32219818234443665, + 0.09207029640674591 + ], + "credit_map": [ + 0.0036102605517953634, + 0.0032848105765879154, + 0.0059738242998719215, + 0.005054619163274765, + 0.004791295621544123, + 0.005679573398083448, + 0.008414391428232193, + 0.009645056910812855, + 0.0023220821749418974, + 0.002875140169635415, + 0.0020498433150351048, + 0.0025940861087292433, + 0.005628418177366257, + 0.004564069211483002, + 0.01039869710803032, + 0.011931400746107101, + 0.003908760845661163, + 0.003212870331481099, + 0.004999678116291761, + 0.008852670900523663, + 0.0041705588810145855, + 0.008501089178025723, + 0.053329020738601685, + 0.04893629625439644, + 0.005932736676186323, + 0.012082269415259361, + 0.006230286322534084, + 0.006154031492769718, + 0.00888335332274437, + 0.011796554550528526, + 0.012504100799560547, + 0.01089865155518055, + 0.002275780076161027, + 0.00281181070022285, + 0.0023808912374079227, + 0.002500792033970356, + 0.0062736584804952145, + 0.0053793746046721935, + 0.003565659513697028, + 0.004325034096837044, + 0.004785533528774977, + 0.004005931317806244, + 0.00444840919226408, + 0.0072482372634112835, + 0.009919765405356884, + 0.008814849890768528, + 0.006455233320593834, + 0.005530260968953371, + 0.006899649277329445, + 0.006110256537795067, + 0.005382895935326815, + 0.0040800259448587894, + 0.04911983013153076, + 0.04867759346961975, + 0.005536757875233889, + 0.005586724728345871, + 0.007935074158012867, + 0.009654182940721512, + 0.005927558522671461, + 0.00814069900661707, + 0.011694652028381824, + 0.011612649075686932, + 0.008759177289903164, + 0.010682053864002228, + 0.04603651165962219, + 0.030892152339220047, + 0.026592088863253593, + 0.028163863345980644, + 0.03862612694501877, + 0.049437880516052246, + 0.05461474880576134, + 0.04783479496836662, + 0.012324432842433453, + 0.007389635778963566, + 0.009252244606614113, + 0.008811437524855137, + 0.009558538906276226, + 0.010816087946295738, + 0.01791183277964592, + 0.01600607857108116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.17992201447486877 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287957534193993 + ], + "opposite_logprob_tokens": [ + -3.7732880115509033 + ], + "logprob_reward": -0.023287957534193993, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287957534193993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0654204711318016, + 0.056246668100357056, + 0.16313208639621735, + 0.11712808907032013, + 0.04281638562679291, + 0.07334890216588974, + 0.08431562036275864, + 0.04554066061973572, + 0.08864618837833405, + 0.26340484619140625 + ], + "credit_map": [ + 0.00493802223354578, + 0.005190239287912846, + 0.006138504482805729, + 0.009399907663464546, + 0.006332383956760168, + 0.008877021260559559, + 0.012204502709209919, + 0.012339891865849495, + 0.0037115735467523336, + 0.005095821805298328, + 0.004325440619140863, + 0.004677006043493748, + 0.0068827723152935505, + 0.005725839640945196, + 0.013143983669579029, + 0.0126842325553298, + 0.007604666519910097, + 0.009564039297401905, + 0.01331162266433239, + 0.03251652047038078, + 0.011980651877820492, + 0.017064908519387245, + 0.03398938849568367, + 0.037100281566381454, + 0.009029778651893139, + 0.020102940499782562, + 0.009844472631812096, + 0.011665074154734612, + 0.01660401001572609, + 0.016246819868683815, + 0.01720898039638996, + 0.016426019370555878, + 0.004223701078444719, + 0.004911972209811211, + 0.004427516832947731, + 0.004828931763768196, + 0.006720827892422676, + 0.0068378085270524025, + 0.006003274116665125, + 0.004862355533987284, + 0.010608009062707424, + 0.0061374809592962265, + 0.005414487328380346, + 0.007800398860126734, + 0.016306409612298012, + 0.014653024263679981, + 0.007263796869665384, + 0.005165295675396919, + 0.007417706307023764, + 0.004841222893446684, + 0.00448395824059844, + 0.004198323003947735, + 0.026306400075554848, + 0.02708445116877556, + 0.005138603504747152, + 0.004844952374696732, + 0.004127562511712313, + 0.004412505775690079, + 0.0035986199509352446, + 0.0038458690978586674, + 0.011613777838647366, + 0.010026821866631508, + 0.004054405726492405, + 0.003861098550260067, + 0.013577754609286785, + 0.008726771920919418, + 0.007410579361021519, + 0.008157300762832165, + 0.009202338755130768, + 0.012590848840773106, + 0.015249093063175678, + 0.013731496408581734, + 0.02753569930791855, + 0.017058713361620903, + 0.02644101344048977, + 0.025579487904906273, + 0.028926247730851173, + 0.029511133208870888, + 0.06138120964169502, + 0.04697134345769882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.023287957534193993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium side shot, rendered as a hyper-realistic CGI animation, tracks smoothly alongside the massive Atlantean city-ship as it descends majestically into the conformist black void of the ocean abyss. The camera maintains a steady, parallel path, tilting slowly downward to emphasize the city's immense scale and its steady, inexorable downward momentum. The city's architecture is a breathtaking fusion of organic forms and advanced technology. Towering crystalline spires shimmer with a soft internal bioluminescence, casting an ethereal divine light of turquoise and gold into the surrounding darkness. Waterfalls appear to originate from within the city, cascading silently down steps, streams catching the inner light and shimmering like liquid diamonds. In stunning, physically accurate detail, the dark waters surround the city's massive, streamlined hull, creating a visible shell of compressed, supercavitated space, seething with microscopic bubbles, while the flow along its flanks appears perfectly laminar and smooth. The scene is filled with an atmosphere of silent awe, capturing a moment of impossible grandeur and ancient mystery with lifelike precision, as clouds of illuminated marine snow drift past like a slow-motion star field.", + "seed": 20497, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_15/index_3713_seed_20497_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.038006126112304625, + "vlm_logprob": -0.038006126112304625, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.038006126112304625, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003538420656695962 + ], + "opposite_logprob_tokens": [ + -6.128538608551025 + ], + "logprob_reward": -0.003538420656695962, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.003538420656695962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32327961921691895, + 0.03616739436984062, + 0.032124727964401245, + 0.03373193368315697, + 0.03748932480812073, + 0.07929953187704086, + 0.20297527313232422, + 0.06382972747087479, + 0.12458735704421997, + 0.06651513278484344 + ], + "credit_map": [ + 0.046821434050798416, + 0.016966624185442924, + 0.13170276582241058, + 0.017348762601614, + 0.01199810579419136, + 0.010908365249633789, + 0.03957894816994667, + 0.04795461520552635, + 0.004783843643963337, + 0.003231092356145382, + 0.003869472537189722, + 0.0034524251241236925, + 0.005782142281532288, + 0.005303156096488237, + 0.0042562526650726795, + 0.005489009898155928, + 0.003795782569795847, + 0.0029374018777161837, + 0.004235736560076475, + 0.004645322449505329, + 0.004130945540964603, + 0.004026962909847498, + 0.004281025379896164, + 0.004071556963026524, + 0.0038134180940687656, + 0.003116417909041047, + 0.0065737441182136536, + 0.0034257671795785427, + 0.003458556719124317, + 0.005377954337745905, + 0.004333333112299442, + 0.003632743377238512, + 0.004410246387124062, + 0.003959469962865114, + 0.004934967495501041, + 0.004236138425767422, + 0.006467064842581749, + 0.005130099132657051, + 0.0044180117547512054, + 0.003933328203856945, + 0.006844849791377783, + 0.006578232627362013, + 0.017018411308526993, + 0.01292747724801302, + 0.00687452033162117, + 0.013236524537205696, + 0.009499668143689632, + 0.006319843232631683, + 0.012760896235704422, + 0.007148271426558495, + 0.06679538637399673, + 0.012020216323435307, + 0.0161299891769886, + 0.01467379555106163, + 0.014680472202599049, + 0.058766260743141174, + 0.009505555965006351, + 0.006797457579523325, + 0.009827736765146255, + 0.007428292650729418, + 0.006227594800293446, + 0.008868830278515816, + 0.008145077154040337, + 0.007029189262539148, + 0.013136863708496094, + 0.012519299983978271, + 0.01728316769003868, + 0.02317461557686329, + 0.011426964774727821, + 0.01157436054199934, + 0.021789979189634323, + 0.013682108372449875, + 0.0073444717563688755, + 0.006906596478074789, + 0.011983857490122318, + 0.0076795644126832485, + 0.0060119968838989735, + 0.013552875258028507, + 0.0062653725035488605, + 0.0067703984677791595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.003538420656695962 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010511117288842797 + ], + "opposite_logprob_tokens": [ + -6.87605094909668 + ], + "logprob_reward": -0.0010511117288842797, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010511117288842797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46241435408592224, + 0.048974137753248215, + 0.03421185538172722, + 0.03315024450421333, + 0.026226237416267395, + 0.04436526820063591, + 0.18319672346115112, + 0.05788162723183632, + 0.0559416301548481, + 0.05363792926073074 + ], + "credit_map": [ + 0.06782499700784683, + 0.024287980049848557, + 0.18505562841892242, + 0.02555682510137558, + 0.017840014770627022, + 0.016987156122922897, + 0.049027182161808014, + 0.07583456486463547, + 0.007119329180568457, + 0.004176918417215347, + 0.007930942811071873, + 0.005239680875092745, + 0.006243360694497824, + 0.006606212817132473, + 0.005930474493652582, + 0.00572721753269434, + 0.0038420320488512516, + 0.0024934292305260897, + 0.005136638879776001, + 0.004907501861453056, + 0.0034002766478806734, + 0.0044043539091944695, + 0.005815205164253712, + 0.004212419502437115, + 0.004582948982715607, + 0.004397880751639605, + 0.006231560371816158, + 0.0030749081633985043, + 0.0032364146318286657, + 0.00515138590708375, + 0.0033252357970923185, + 0.003149912226945162, + 0.0028898639138787985, + 0.002792860846966505, + 0.0030812090262770653, + 0.0038900456856936216, + 0.003750842995941639, + 0.0032723506446927786, + 0.003623064374551177, + 0.0029259975999593735, + 0.0037111137062311172, + 0.003045819466933608, + 0.006529804784804583, + 0.008855181746184826, + 0.004479639232158661, + 0.005759640596807003, + 0.007008951622992754, + 0.004975117743015289, + 0.012169473804533482, + 0.006270561832934618, + 0.061959054321050644, + 0.011358114890754223, + 0.013181541115045547, + 0.013468168675899506, + 0.011344484984874725, + 0.05344533175230026, + 0.007036986295133829, + 0.005280422046780586, + 0.011803057044744492, + 0.005618583876639605, + 0.004954139236360788, + 0.004827637225389481, + 0.006528398022055626, + 0.01183240208774805, + 0.006979511119425297, + 0.004832128528505564, + 0.008379548788070679, + 0.011532481759786606, + 0.004944991786032915, + 0.004560999106615782, + 0.008990099653601646, + 0.0057218726724386215, + 0.005509456153959036, + 0.00549561670050025, + 0.008921394124627113, + 0.005822181235998869, + 0.004927236121147871, + 0.009863707236945629, + 0.00545786926522851, + 0.007640467956662178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0010511117288842797 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34698420763015747, + 0.06411322951316833, + 0.07040127366781235, + 0.037867989391088486, + 0.07757623493671417, + 0.06689362972974777, + 0.128773033618927, + 0.03895137086510658, + 0.08608034253120422, + 0.0823586955666542 + ], + "credit_map": [ + 0.05395420268177986, + 0.018076932057738304, + 0.12867175042629242, + 0.01776883937418461, + 0.012557204812765121, + 0.012793055735528469, + 0.04836580529808998, + 0.054796405136585236, + 0.008112010546028614, + 0.004802484530955553, + 0.007421290967613459, + 0.007637692615389824, + 0.010116039775311947, + 0.006891075987368822, + 0.009932637214660645, + 0.009199993684887886, + 0.009241202846169472, + 0.00676498981192708, + 0.00947690848261118, + 0.0074900174513459206, + 0.007182927802205086, + 0.009537730365991592, + 0.011481394991278648, + 0.009226096794009209, + 0.005075836554169655, + 0.00412164768204093, + 0.008154191076755524, + 0.003788369707763195, + 0.0034895550925284624, + 0.004908017814159393, + 0.0037853482645004988, + 0.004545022267848253, + 0.009620612487196922, + 0.007265944499522448, + 0.009456772357225418, + 0.010663977824151516, + 0.008988883346319199, + 0.011487607844173908, + 0.00911650899797678, + 0.010975917801260948, + 0.005356288980692625, + 0.00432620570063591, + 0.010284307412803173, + 0.014482627622783184, + 0.0062285722233355045, + 0.00726746441796422, + 0.011663299053907394, + 0.007284858729690313, + 0.008480236865580082, + 0.00489769084379077, + 0.04130891337990761, + 0.008480492047965527, + 0.009284361265599728, + 0.009679646231234074, + 0.009629352949559689, + 0.037012338638305664, + 0.0062980917282402515, + 0.003343252930790186, + 0.007996873930096626, + 0.0037662340328097343, + 0.0027473296504467726, + 0.003513720817863941, + 0.004594720434397459, + 0.006691142916679382, + 0.009172395803034306, + 0.008066384121775627, + 0.012339677661657333, + 0.017315054312348366, + 0.007651364430785179, + 0.008294746279716492, + 0.013191481120884418, + 0.010049226693809032, + 0.01080342661589384, + 0.011068113148212433, + 0.016507823020219803, + 0.008802554570138454, + 0.006498835515230894, + 0.014583765529096127, + 0.0057327719405293465, + 0.00836139265447855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.012536413036286831 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036780178546906 + ], + "opposite_logprob_tokens": [ + -1.9103678464889526 + ], + "logprob_reward": -0.16036780178546906, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036780178546906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4395757019519806, + 0.053968269377946854, + 0.041989799588918686, + 0.04946355149149895, + 0.09763111174106598, + 0.055988434702157974, + 0.06524542719125748, + 0.046426016837358475, + 0.08501199632883072, + 0.06469962745904922 + ], + "credit_map": [ + 0.07443445920944214, + 0.023472139611840248, + 0.18954245746135712, + 0.02432044968008995, + 0.01422237977385521, + 0.012598355300724506, + 0.05269460380077362, + 0.04829089343547821, + 0.005591586697846651, + 0.003736332757398486, + 0.007473357487469912, + 0.006446580868214369, + 0.006966735702008009, + 0.007961239665746689, + 0.00689690001308918, + 0.008895538747310638, + 0.006154736969619989, + 0.0036835891660302877, + 0.0057172453962266445, + 0.0054917167872190475, + 0.00548383267596364, + 0.0053190370090305805, + 0.004697692114859819, + 0.005441950634121895, + 0.0063858735375106335, + 0.005447077099233866, + 0.010513619519770145, + 0.005849230568856001, + 0.0046979160979390144, + 0.0053812051191926, + 0.00578253623098135, + 0.005406096111983061, + 0.0104133952409029, + 0.013866535387933254, + 0.010548261925578117, + 0.013182740658521652, + 0.014888596720993519, + 0.012743606232106686, + 0.012494983151555061, + 0.009493003599345684, + 0.0056035807356238365, + 0.006428156513720751, + 0.009045131504535675, + 0.008039366453886032, + 0.008688305504620075, + 0.008621662855148315, + 0.005385649390518665, + 0.00417658407241106, + 0.004388731438666582, + 0.004175685346126556, + 0.015180753543972969, + 0.007842189632356167, + 0.006028841249644756, + 0.006561866495758295, + 0.01078005786985159, + 0.01028730534017086, + 0.00727495364844799, + 0.004240233451128006, + 0.008461663499474525, + 0.004257233813405037, + 0.00401098420843482, + 0.005989577155560255, + 0.005611462518572807, + 0.006579914130270481, + 0.008493316359817982, + 0.007603603880852461, + 0.013430002145469189, + 0.016939634457230568, + 0.006679263897240162, + 0.007929585874080658, + 0.013572708703577518, + 0.010363882407546043, + 0.006931498181074858, + 0.009382170625030994, + 0.014858155511319637, + 0.007724810857325792, + 0.006538787391036749, + 0.008004617877304554, + 0.0057275258004665375, + 0.005532070528715849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.16036780178546906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536883354187012 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012536883354187012, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536883354187012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5157671570777893, + 0.03400018811225891, + 0.014590702950954437, + 0.04214119538664818, + 0.03718365728855133, + 0.026585714891552925, + 0.03391259163618088, + 0.04348708316683769, + 0.12393565475940704, + 0.12839609384536743 + ], + "credit_map": [ + 0.07972052693367004, + 0.026464423164725304, + 0.19787999987602234, + 0.028829995542764664, + 0.01604611799120903, + 0.021764906123280525, + 0.03965169936418533, + 0.10540945827960968, + 0.0040881033055484295, + 0.0025411054957658052, + 0.004107609856873751, + 0.00655363779515028, + 0.0042623369954526424, + 0.0030078087002038956, + 0.005194077733904123, + 0.004245510790497065, + 0.001991637982428074, + 0.001566106453537941, + 0.002072148723527789, + 0.0020121296402066946, + 0.0012710440205410123, + 0.0018143816851079464, + 0.0017734852153807878, + 0.002089770045131445, + 0.006354234647005796, + 0.0045665353536605835, + 0.006153457798063755, + 0.004486903548240662, + 0.004096326883882284, + 0.007532683666795492, + 0.005146700423210859, + 0.003804351668804884, + 0.003991442732512951, + 0.004108182154595852, + 0.004509142599999905, + 0.005136440042406321, + 0.005836407653987408, + 0.004809828475117683, + 0.004284250549972057, + 0.004507965408265591, + 0.003318405942991376, + 0.0022371115628629923, + 0.004666289314627647, + 0.003540985519066453, + 0.0027546475175768137, + 0.004281340632587671, + 0.002382506849244237, + 0.0034044261556118727, + 0.003282250603660941, + 0.00270268926396966, + 0.006842016242444515, + 0.004316400736570358, + 0.0028190696612000465, + 0.0034930792171508074, + 0.005399416666477919, + 0.005057670176029205, + 0.007367964833974838, + 0.005617935210466385, + 0.006617442704737186, + 0.005007038824260235, + 0.00546626839786768, + 0.0040476140566170216, + 0.004759860225021839, + 0.004602956585586071, + 0.013187164440751076, + 0.010046832263469696, + 0.02059049904346466, + 0.02110082283616066, + 0.011861220002174377, + 0.012527703307569027, + 0.01776534877717495, + 0.01685607060790062, + 0.012471293099224567, + 0.020673664286732674, + 0.020443789660930634, + 0.017385508865118027, + 0.01589863747358322, + 0.017473354935646057, + 0.01256589312106371, + 0.011483961716294289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.012536883354187012 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium side shot, rendered as a hyper-realistic CGI animation, tracks smoothly alongside the massive Atlantean city-ship as it descends majestically into the conformist black void of the ocean abyss. The camera maintains a steady, parallel path, tilting slowly downward to emphasize the city's immense scale and its steady, inexorable downward momentum. The city's architecture is a breathtaking fusion of organic forms and advanced technology. Towering crystalline spires shimmer with a soft internal bioluminescence, casting an ethereal divine light of turquoise and gold into the surrounding darkness. Waterfalls appear to originate from within the city, cascading silently down steps, streams catching the inner light and shimmering like liquid diamonds. In stunning, physically accurate detail, the dark waters surround the city's massive, streamlined hull, creating a visible shell of compressed, supercavitated space, seething with microscopic bubbles, while the flow along its flanks appears perfectly laminar and smooth. The scene is filled with an atmosphere of silent awe, capturing a moment of impossible grandeur and ancient mystery with lifelike precision, as clouds of illuminated marine snow drift past like a slow-motion star field.", + "seed": 20498, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_15/index_3713_seed_20498_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.019480620324611665, + "vlm_logprob": -0.019480620324611665, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.019480620324611665, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036566078197211027 + ], + "opposite_logprob_tokens": [ + -6.253656387329102 + ], + "logprob_reward": -0.0036566078197211027, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0036566078197211027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38470280170440674, + 0.02454247698187828, + 0.1313973367214203, + 0.03915202245116234, + 0.05649111792445183, + 0.02333837002515793, + 0.054402031004428864, + 0.06685810536146164, + 0.1477639079093933, + 0.07135190069675446 + ], + "credit_map": [ + 0.02936803735792637, + 0.016570523381233215, + 0.1390794962644577, + 0.03702981397509575, + 0.023333588615059853, + 0.0949079766869545, + 0.022753847762942314, + 0.021659543737769127, + 0.004564498085528612, + 0.0032456472981721163, + 0.002501880517229438, + 0.002169250510632992, + 0.0034061584156006575, + 0.003925435710698366, + 0.002314415294677019, + 0.002415191847831011, + 0.01115439273416996, + 0.025624386966228485, + 0.016395583748817444, + 0.01224418357014656, + 0.01771550625562668, + 0.018889257684350014, + 0.013781446032226086, + 0.015592571347951889, + 0.002991746412590146, + 0.003167189424857497, + 0.004451450891792774, + 0.00656277546659112, + 0.0024623603094369173, + 0.007775445934385061, + 0.007203695364296436, + 0.0045373584143817425, + 0.003646447090432048, + 0.0036747476551681757, + 0.01662627048790455, + 0.006216120440512896, + 0.003139892127364874, + 0.004584669601172209, + 0.0037929050158709288, + 0.014810064807534218, + 0.0022100370842963457, + 0.0031757066026329994, + 0.0035133317578583956, + 0.002737596398219466, + 0.0024525721091777086, + 0.0026309676468372345, + 0.0032520736567676067, + 0.0033660843037068844, + 0.003709424752742052, + 0.009208000265061855, + 0.007630614563822746, + 0.0075233145616948605, + 0.006550427991896868, + 0.0059164720587432384, + 0.008133304305374622, + 0.005730475299060345, + 0.005414854735136032, + 0.00576329929754138, + 0.008846520446240902, + 0.01143618393689394, + 0.009517746977508068, + 0.009461671113967896, + 0.010149206034839153, + 0.006268620025366545, + 0.008436726406216621, + 0.008054696023464203, + 0.016722597181797028, + 0.0248058270663023, + 0.015407278202474117, + 0.023540914058685303, + 0.027317285537719727, + 0.023478573188185692, + 0.0045682811178267, + 0.006647126749157906, + 0.009409988299012184, + 0.012134851887822151, + 0.009928781539201736, + 0.011893953196704388, + 0.010369009338319302, + 0.006399902049452066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0036566078197211027 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017219253350049257 + ], + "opposite_logprob_tokens": [ + -6.3767218589782715 + ], + "logprob_reward": -0.0017219253350049257, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017219253350049257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4428449869155884, + 0.025637349113821983, + 0.05637158453464508, + 0.04226941615343094, + 0.1708505004644394, + 0.020945357158780098, + 0.032166387885808945, + 0.03423355147242546, + 0.08247577399015427, + 0.09220508486032486 + ], + "credit_map": [ + 0.03469076380133629, + 0.017801256850361824, + 0.17216689884662628, + 0.03527018055319786, + 0.023736316710710526, + 0.11987908184528351, + 0.02263473905622959, + 0.016665760427713394, + 0.0032017589546740055, + 0.0036875701043754816, + 0.003920705523341894, + 0.003118728520348668, + 0.0021395201329141855, + 0.00400103535503149, + 0.002344151260331273, + 0.0032238797284662724, + 0.006856885273009539, + 0.009052188135683537, + 0.007231927011162043, + 0.005239332560449839, + 0.006657525431364775, + 0.007077440619468689, + 0.006781740114092827, + 0.007474543526768684, + 0.006006250157952309, + 0.004690608009696007, + 0.004229014273732901, + 0.005259169731289148, + 0.0032296564895659685, + 0.006865155883133411, + 0.006891926284879446, + 0.005097635090351105, + 0.008887908421456814, + 0.008608941920101643, + 0.060465048998594284, + 0.01198333129286766, + 0.0055013615638017654, + 0.008437246084213257, + 0.006441583391278982, + 0.060525067150592804, + 0.001489652437157929, + 0.002377243246883154, + 0.003329712897539139, + 0.0025032993871718645, + 0.002191385021433234, + 0.002643121872097254, + 0.003347323276102543, + 0.0030636191368103027, + 0.0028194759506732225, + 0.004812190309166908, + 0.0044509428553283215, + 0.0038990096654742956, + 0.003725293558090925, + 0.004382733255624771, + 0.004644700326025486, + 0.003432042431086302, + 0.0031701496336609125, + 0.0025097623001784086, + 0.004568381700664759, + 0.005310705862939358, + 0.004437700845301151, + 0.004618501290678978, + 0.0063529470935463905, + 0.0032654039096087217, + 0.005693336948752403, + 0.005256286822259426, + 0.01522687915712595, + 0.017109721899032593, + 0.010320605710148811, + 0.012618412263691425, + 0.009061845950782299, + 0.007188679650425911, + 0.0044536045752465725, + 0.006402176804840565, + 0.007640280295163393, + 0.006111724302172661, + 0.01865195482969284, + 0.02995315007865429, + 0.0062414309941232204, + 0.012750759720802307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0017219253350049257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065318137407303 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016065318137407303, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065318137407303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4710090756416321, + 0.04673660919070244, + 0.07032005488872528, + 0.044125035405159, + 0.1340598165988922, + 0.04314063489437103, + 0.07810834795236588, + 0.02446623519062996, + 0.047966234385967255, + 0.040067967027425766 + ], + "credit_map": [ + 0.03633740544319153, + 0.014553201384842396, + 0.18566668033599854, + 0.039013274013996124, + 0.022729109972715378, + 0.12895196676254272, + 0.02412412129342556, + 0.01963331177830696, + 0.00608432712033391, + 0.0062329270876944065, + 0.004475078545510769, + 0.0047481670044362545, + 0.006904239300638437, + 0.0067914268001914024, + 0.004460201598703861, + 0.007040239404886961, + 0.007151575293391943, + 0.013075570575892925, + 0.008599734865128994, + 0.007246225606650114, + 0.01096484437584877, + 0.008853464387357235, + 0.006279315333813429, + 0.008149323984980583, + 0.0045314086601138115, + 0.004129623994231224, + 0.005080505274236202, + 0.006226981990039349, + 0.00313545111566782, + 0.008822857402265072, + 0.007209109608083963, + 0.004989094100892544, + 0.010558650828897953, + 0.012925890274345875, + 0.0388256274163723, + 0.013875843025743961, + 0.007410550490021706, + 0.00815117172896862, + 0.006593658588826656, + 0.035718437284231186, + 0.003941715694963932, + 0.004713514819741249, + 0.008090363815426826, + 0.005387665703892708, + 0.004068650305271149, + 0.0039786952547729015, + 0.005506575107574463, + 0.007453454192727804, + 0.004850675351917744, + 0.008661741390824318, + 0.01454914640635252, + 0.01045137457549572, + 0.009838469326496124, + 0.009674092754721642, + 0.011205286718904972, + 0.008877556771039963, + 0.002422217046841979, + 0.0020467606373131275, + 0.003353018080815673, + 0.003843599697574973, + 0.0028494358994066715, + 0.004273987375199795, + 0.0031391156371682882, + 0.0025380998849868774, + 0.003045442281290889, + 0.0026702582836151123, + 0.0065909819677472115, + 0.011712955310940742, + 0.00595445279031992, + 0.007706293370574713, + 0.0049706329591572285, + 0.005315216723829508, + 0.0016809436492621899, + 0.0019362698076292872, + 0.003426519688218832, + 0.004458772484213114, + 0.007348678540438414, + 0.012390128336846828, + 0.0031748537439852953, + 0.005651801358908415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.016065318137407303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001114636659622 + ], + "opposite_logprob_tokens": [ + -2.6950111389160156 + ], + "logprob_reward": -0.07001114636659622, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001114636659622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3809926211833954, + 0.028630251064896584, + 0.11316772550344467, + 0.039542458951473236, + 0.201790452003479, + 0.03533562272787094, + 0.06457056105136871, + 0.033502861857414246, + 0.0640704557299614, + 0.03839690238237381 + ], + "credit_map": [ + 0.028808211907744408, + 0.010969143360853195, + 0.15370306372642517, + 0.031031668186187744, + 0.017457878217101097, + 0.10607787221670151, + 0.018224023282527924, + 0.014720771461725235, + 0.004461127333343029, + 0.0030335637275129557, + 0.0025301072746515274, + 0.002884474815800786, + 0.0040263826958835125, + 0.004446140490472317, + 0.003323496552184224, + 0.003924959804862738, + 0.009978443384170532, + 0.015180583111941814, + 0.014354744926095009, + 0.012635243125259876, + 0.010550389997661114, + 0.01158022042363882, + 0.022770801559090614, + 0.016117312014102936, + 0.003965381998568773, + 0.0030623364727944136, + 0.004917443264275789, + 0.00507838511839509, + 0.0029061667155474424, + 0.008411056362092495, + 0.006338948383927345, + 0.004862741567194462, + 0.008795366622507572, + 0.009552599862217903, + 0.07311592996120453, + 0.013855143450200558, + 0.005791964940726757, + 0.00971096009016037, + 0.0077359057031571865, + 0.07323259860277176, + 0.002710326574742794, + 0.0032801202032715082, + 0.004622745327651501, + 0.0056421272456645966, + 0.003991094417870045, + 0.005383949726819992, + 0.005333870183676481, + 0.004371390677988529, + 0.003968863748013973, + 0.009069832041859627, + 0.011243629269301891, + 0.00993589498102665, + 0.012138578109443188, + 0.005863029044121504, + 0.006119471974670887, + 0.006231263745576143, + 0.002396490192040801, + 0.003322772216051817, + 0.00504272012040019, + 0.005815302953124046, + 0.006002385169267654, + 0.00416378490626812, + 0.003966346848756075, + 0.002793064573779702, + 0.003660933580249548, + 0.004604706075042486, + 0.00879824161529541, + 0.010001527145504951, + 0.006625944748520851, + 0.008122850209474564, + 0.011485181748867035, + 0.010771075263619423, + 0.0018375604413449764, + 0.0031411899253726006, + 0.004240077454596758, + 0.005329377017915249, + 0.005553883500397205, + 0.008008326403796673, + 0.005648014135658741, + 0.004638475365936756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.07001114636659622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive Atlantean city-ship with organic and technological architecture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the city's spires crystalline and emitting turquoise and gold bioluminescent light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do waterfalls originate from within the city and cascade down its steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a visible shell of compressed, supercavitated space with microscopic bubbles surrounding the city's hull?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera track smoothly alongside the ship while tilting slowly downward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948103964328766 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005948103964328766, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948103964328766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2957620620727539, + 0.050684500485658646, + 0.06830397993326187, + 0.06300803273916245, + 0.17369453608989716, + 0.10606774687767029, + 0.03945751115679741, + 0.059720348566770554, + 0.09213987737894058, + 0.05116141960024834 + ], + "credit_map": [ + 0.02230040170252323, + 0.009288365952670574, + 0.1183963268995285, + 0.02508500963449478, + 0.013808047398924828, + 0.08150779455900192, + 0.014177275821566582, + 0.011198843829333782, + 0.008128546178340912, + 0.00691945431753993, + 0.00451077613979578, + 0.004550638142973185, + 0.007713903207331896, + 0.007561694364994764, + 0.004268863704055548, + 0.007030623499304056, + 0.007814904674887657, + 0.008153553120791912, + 0.010121352039277554, + 0.008460582233965397, + 0.007434917148202658, + 0.007819361984729767, + 0.008689677342772484, + 0.009809635579586029, + 0.008095864206552505, + 0.007640610449016094, + 0.005995132494717836, + 0.008561372756958008, + 0.005140568595379591, + 0.008901956491172314, + 0.010126038454473019, + 0.008546489290893078, + 0.008195017464458942, + 0.007797540165483952, + 0.061350513249635696, + 0.011887066066265106, + 0.005657567176967859, + 0.008951696567237377, + 0.0080042639747262, + 0.06185086816549301, + 0.007641398347914219, + 0.008347722701728344, + 0.01690766029059887, + 0.01268014870584011, + 0.011524911969900131, + 0.015566890127956867, + 0.019823584705591202, + 0.013575433753430843, + 0.0035976171493530273, + 0.005876528099179268, + 0.007264875806868076, + 0.0058989073149859905, + 0.003985918592661619, + 0.00396799948066473, + 0.005069281905889511, + 0.003796382574364543, + 0.004567904397845268, + 0.005120655056089163, + 0.007475509773939848, + 0.011379792355000973, + 0.007256018463522196, + 0.006453592795878649, + 0.011995612643659115, + 0.005471261218190193, + 0.004624749533832073, + 0.004182897973805666, + 0.014132233336567879, + 0.010999929159879684, + 0.011596712283790112, + 0.015238738618791103, + 0.018825193867087364, + 0.012539421208202839, + 0.0021074197720736265, + 0.0033487447071820498, + 0.00428928155452013, + 0.0037402950692921877, + 0.008893649093806744, + 0.01698518916964531, + 0.0035465899854898453, + 0.008250247687101364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.005948103964328766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_2/index_tensor([1145, 1145])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_2/index_tensor([1145, 1145])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7901631c0bf90da8b97545dabf821a9c3974c1bf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_2/index_tensor([1145, 1145])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic low-angle, wide-angle shot captures the scene, the camera initially static to establish the monumental scale, but then slowly, deliberately tilting upward to follow the unfolding action. At dusk, beneath a turbulent sky of deep purple and scarred orange clouds, Paris's iconic iron lattice tower stands tall, its intricate cast-iron latticework rendered in sharp, lifelike detail. Suddenly, a violent and profound tremor sweeps through the vast structure, causing dust and small metal fragments to break free from its beams. The tower's base is instantly engulfed in a roaring inferno of brilliant orange and yellow flames, releasing a vast, billowing cloud of thick, black and gray smoke that obscures the ground. With the immense power of a massive rocket, the entire tower begins to launch skyward in a powerful and relentless ascent, steadily defying gravity. Shockwaves of flame and debris erupt from the tower's foundation, creating a surreal and terrifying atmosphere of destruction.", + "seed": 15287, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_2/index_1145_seed_15287_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0474748631939292, + "vlm_logprob": -0.0474748631939292, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.0474748631939292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05037752166390419 + ], + "opposite_logprob_tokens": [ + -3.050377607345581 + ], + "logprob_reward": -0.05037752166390419, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05037752166390419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08165343105792999, + 0.05351919308304787, + 0.07643978297710419, + 0.07482805103063583, + 0.20118661224842072, + 0.0535416342318058, + 0.08105301856994629, + 0.07221114635467529, + 0.13929162919521332, + 0.1662754863500595 + ], + "credit_map": [ + 0.010413550771772861, + 0.0067139845341444016, + 0.010583307594060898, + 0.013473388738930225, + 0.00879356637597084, + 0.011541349813342094, + 0.011069376021623611, + 0.009064906276762486, + 0.00552710983902216, + 0.0067670331336557865, + 0.006449621170759201, + 0.005467974580824375, + 0.004742744844406843, + 0.008195366710424423, + 0.0047180731780827045, + 0.011651270091533661, + 0.011658713221549988, + 0.0075414422899484634, + 0.010739102028310299, + 0.006217691116034985, + 0.0065645636059343815, + 0.01236556377261877, + 0.0071139466017484665, + 0.014238759875297546, + 0.009155705571174622, + 0.013056635856628418, + 0.008913636207580566, + 0.012004737742245197, + 0.005923132877796888, + 0.006656236480921507, + 0.010634795762598515, + 0.008483176119625568, + 0.021243086084723473, + 0.025870488956570625, + 0.02913334220647812, + 0.028273288160562515, + 0.02516017109155655, + 0.03033844195306301, + 0.021717272698879242, + 0.019450534135103226, + 0.0112300468608737, + 0.007739244494587183, + 0.01039210893213749, + 0.006638643331825733, + 0.004883951507508755, + 0.0042223031632602215, + 0.004766729660332203, + 0.003668606746941805, + 0.01262671034783125, + 0.01444527879357338, + 0.009132837876677513, + 0.0077376109547913074, + 0.008543242700397968, + 0.007646519690752029, + 0.011863953433930874, + 0.00905686803162098, + 0.01920698583126068, + 0.013312753289937973, + 0.012528755702078342, + 0.00605944637209177, + 0.0066881016828119755, + 0.005358131602406502, + 0.004859162960201502, + 0.004197807051241398, + 0.019550316035747528, + 0.029494525864720345, + 0.014516545459628105, + 0.0075084008276462555, + 0.01277556549757719, + 0.03458377718925476, + 0.012457002885639668, + 0.00840549822896719, + 0.023564942181110382, + 0.023631324991583824, + 0.03575356677174568, + 0.01464114896953106, + 0.027562586590647697, + 0.017803024500608444, + 0.015563640743494034, + 0.007755249738693237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.05037752166390419 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007650598883628845 + ], + "opposite_logprob_tokens": [ + -4.882650375366211 + ], + "logprob_reward": -0.007650598883628845, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007650598883628845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10913804173469543, + 0.08911623060703278, + 0.04648350179195404, + 0.06413242220878601, + 0.10756826400756836, + 0.0757957473397255, + 0.07618731260299683, + 0.10684197396039963, + 0.19999846816062927, + 0.1247379407286644 + ], + "credit_map": [ + 0.021428372710943222, + 0.009797410108149052, + 0.016370613127946854, + 0.01196880079805851, + 0.009390408173203468, + 0.015696343034505844, + 0.009301629848778248, + 0.015184465795755386, + 0.013147402554750443, + 0.009530488401651382, + 0.012016315944492817, + 0.010935223661363125, + 0.007028196007013321, + 0.011684325523674488, + 0.008602986112236977, + 0.016171304509043694, + 0.005509784910827875, + 0.004350683651864529, + 0.007171437609940767, + 0.003737484570592642, + 0.004578690975904465, + 0.006132515612989664, + 0.006715930067002773, + 0.008286979049444199, + 0.010940869338810444, + 0.008787070401012897, + 0.010244754143059254, + 0.010761018842458725, + 0.003995505161583424, + 0.006121260579675436, + 0.006807571277022362, + 0.006474378984421492, + 0.016433075070381165, + 0.013526021502912045, + 0.023623889312148094, + 0.010637265630066395, + 0.012008695863187313, + 0.013524449430406094, + 0.0099985646083951, + 0.007816320285201073, + 0.01659286580979824, + 0.013530085794627666, + 0.02222391776740551, + 0.007170424330979586, + 0.004016381688416004, + 0.0037420885637402534, + 0.0047980076633393764, + 0.003721986198797822, + 0.01534378994256258, + 0.015006421133875847, + 0.017272034659981728, + 0.005349648185074329, + 0.005409052595496178, + 0.005430807825177908, + 0.006543992552906275, + 0.005831574089825153, + 0.029091114178299904, + 0.020504023879766464, + 0.02248009666800499, + 0.008137699216604233, + 0.0075351642444729805, + 0.00728280795738101, + 0.006821464281529188, + 0.004989621229469776, + 0.028460852801799774, + 0.029999621212482452, + 0.027850575745105743, + 0.020990785211324692, + 0.022618835791945457, + 0.03079071268439293, + 0.02304087020456791, + 0.016246221959590912, + 0.017269285395741463, + 0.01321751531213522, + 0.023161407560110092, + 0.018709639087319374, + 0.01663750782608986, + 0.013587323017418385, + 0.01340517308562994, + 0.008750112727284431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.007650598883628845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420364435762167 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.01420364435762167, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420364435762167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08355507999658585, + 0.023926349356770515, + 0.045229192823171616, + 0.06151187792420387, + 0.04076646268367767, + 0.04612685367465019, + 0.2637629508972168, + 0.1399531364440918, + 0.06961804628372192, + 0.2255500853061676 + ], + "credit_map": [ + 0.01063918974250555, + 0.006375932600349188, + 0.014999178238213062, + 0.01038367673754692, + 0.01008745189756155, + 0.01158095058053732, + 0.009058156982064247, + 0.010430538095533848, + 0.0035371221601963043, + 0.002402651822194457, + 0.003217751160264015, + 0.002787096193060279, + 0.0027638794854283333, + 0.0036208471283316612, + 0.0022813098039478064, + 0.003315690439194441, + 0.004419350530952215, + 0.0047186268493533134, + 0.005240469239652157, + 0.005555284209549427, + 0.005258533172309399, + 0.006708094850182533, + 0.006399822421371937, + 0.006929007358849049, + 0.007323039695620537, + 0.007822751998901367, + 0.009842225350439548, + 0.007535104174166918, + 0.0048988680355250835, + 0.01009534765034914, + 0.007229367271065712, + 0.006765169557183981, + 0.005051386542618275, + 0.004083100240677595, + 0.005590974818915129, + 0.004958593752235174, + 0.0045648724772036076, + 0.00449353875592351, + 0.006438226904720068, + 0.005585767794400454, + 0.006709463894367218, + 0.007220768369734287, + 0.0062126279808580875, + 0.005365564487874508, + 0.0048584043979644775, + 0.0035621037241071463, + 0.007111827377229929, + 0.005086089950054884, + 0.06060047820210457, + 0.06090737134218216, + 0.05421784520149231, + 0.013652987778186798, + 0.019921500235795975, + 0.019687460735440254, + 0.01966913789510727, + 0.015106146223843098, + 0.04236486926674843, + 0.02667813189327717, + 0.02798331528902054, + 0.010235452093183994, + 0.007897102274000645, + 0.010992384515702724, + 0.008614659309387207, + 0.0051871961914002895, + 0.013332484290003777, + 0.010358965955674648, + 0.013485213741660118, + 0.0065231011249125, + 0.008806044235825539, + 0.0077128284610807896, + 0.0054512289352715015, + 0.00394816929474473, + 0.029886366799473763, + 0.022819675505161285, + 0.04125077277421951, + 0.036422841250896454, + 0.01947149634361267, + 0.022915489971637726, + 0.03154059126973152, + 0.021242840215563774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.01420364435762167 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704616785049438 + ], + "opposite_logprob_tokens": [ + -2.1270461082458496 + ], + "logprob_reward": -0.12704616785049438, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704616785049438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049700040370225906, + 0.04596849903464317, + 0.048174530267715454, + 0.07774762809276581, + 0.06516248732805252, + 0.07048392295837402, + 0.07222054153680801, + 0.2504142224788666, + 0.2100650668144226, + 0.1100630909204483 + ], + "credit_map": [ + 0.004618706181645393, + 0.0052412329241633415, + 0.006649815011769533, + 0.008889058604836464, + 0.005272462964057922, + 0.006998492404818535, + 0.007416146341711283, + 0.00461412500590086, + 0.0053699384443461895, + 0.004253744147717953, + 0.00653136195614934, + 0.0044755516573786736, + 0.0064170463010668755, + 0.006659747567027807, + 0.005010730586946011, + 0.007250380236655474, + 0.005672555882483721, + 0.004609198309481144, + 0.005860465113073587, + 0.005217419937252998, + 0.005320481490343809, + 0.007002341561019421, + 0.0061040641739964485, + 0.008388001471757889, + 0.013033931143581867, + 0.010049655102193356, + 0.010901862755417824, + 0.01412301603704691, + 0.005274686962366104, + 0.006983133498579264, + 0.007104154210537672, + 0.01027719210833311, + 0.006593544501811266, + 0.006121700629591942, + 0.007440329063683748, + 0.010886835865676403, + 0.0066843219101428986, + 0.008385556749999523, + 0.007548082619905472, + 0.011502119712531567, + 0.0150290010496974, + 0.010114619508385658, + 0.015335369855165482, + 0.00821777991950512, + 0.00566477607935667, + 0.004183136858046055, + 0.00638775946572423, + 0.005551480688154697, + 0.0167347751557827, + 0.015551643446087837, + 0.011365548707544804, + 0.0041592963971197605, + 0.007291228976100683, + 0.007221504580229521, + 0.0053815413266420364, + 0.004514994099736214, + 0.06340554356575012, + 0.04739326983690262, + 0.03710480406880379, + 0.018516821786761284, + 0.024173468351364136, + 0.022681206464767456, + 0.020565999671816826, + 0.016573110595345497, + 0.024767925962805748, + 0.02573077380657196, + 0.0348709337413311, + 0.025789566338062286, + 0.03190018609166145, + 0.024196360260248184, + 0.02452852576971054, + 0.018280791118741035, + 0.013970376923680305, + 0.01770523004233837, + 0.02008976601064205, + 0.01000721100717783, + 0.01687091588973999, + 0.011033526621758938, + 0.013323278166353703, + 0.007062786724418402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.12704616785049438 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809638321399689 + ], + "opposite_logprob_tokens": [ + -3.2880964279174805 + ], + "logprob_reward": -0.03809638321399689, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809638321399689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06414356082677841, + 0.035455089062452316, + 0.04940075799822807, + 0.06535491347312927, + 0.05729454383254051, + 0.062082648277282715, + 0.09982278943061829, + 0.15211376547813416, + 0.27909448742866516, + 0.1352374404668808 + ], + "credit_map": [ + 0.009901080280542374, + 0.009172036312520504, + 0.010877986438572407, + 0.006634423974901438, + 0.006130294408649206, + 0.010641866363584995, + 0.004840993322432041, + 0.005944871809333563, + 0.004164247307926416, + 0.0038069530855864286, + 0.006359136197715998, + 0.005678815301507711, + 0.003554467111825943, + 0.005300501827150583, + 0.002590878400951624, + 0.004000086337327957, + 0.005293793976306915, + 0.005611020606011152, + 0.005703663919121027, + 0.0046201348304748535, + 0.00324952881783247, + 0.012342149391770363, + 0.0035816647578030825, + 0.008998798206448555, + 0.009127904660999775, + 0.01251605711877346, + 0.009630030952394009, + 0.014042478054761887, + 0.0034662808757275343, + 0.004698995500802994, + 0.006348478142172098, + 0.0055246842093765736, + 0.00785948894917965, + 0.00528660649433732, + 0.00865776464343071, + 0.009376047179102898, + 0.004933157004415989, + 0.006932906806468964, + 0.007500925101339817, + 0.006747642997652292, + 0.010274133644998074, + 0.008495453745126724, + 0.014262179844081402, + 0.0076654404401779175, + 0.004772996064275503, + 0.004340784624218941, + 0.007550582755357027, + 0.00472107483074069, + 0.023325355723500252, + 0.02113453857600689, + 0.024023929610848427, + 0.005465445574373007, + 0.008062301203608513, + 0.0075086201541125774, + 0.005788455717265606, + 0.004514138679951429, + 0.028564421460032463, + 0.024245403707027435, + 0.025906411930918694, + 0.012883612886071205, + 0.015557060949504375, + 0.021503329277038574, + 0.013601073063910007, + 0.009852441027760506, + 0.03646121546626091, + 0.058122601360082626, + 0.029813941568136215, + 0.021276839077472687, + 0.025501247495412827, + 0.06929142028093338, + 0.021609505638480186, + 0.017017686739563942, + 0.01719796657562256, + 0.018728068098425865, + 0.016959236934781075, + 0.017218448221683502, + 0.027033420279622078, + 0.013977293856441975, + 0.014515149407088757, + 0.00960784126073122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.03809638321399689 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle, wide-angle shot captures the scene, the camera initially static to establish the monumental scale, but then slowly, deliberately tilting upward to follow the unfolding action. At dusk, beneath a turbulent sky of deep purple and scarred orange clouds, Paris's iconic iron lattice tower stands tall, its intricate cast-iron latticework rendered in sharp, lifelike detail. Suddenly, a violent and profound tremor sweeps through the vast structure, causing dust and small metal fragments to break free from its beams. The tower's base is instantly engulfed in a roaring inferno of brilliant orange and yellow flames, releasing a vast, billowing cloud of thick, black and gray smoke that obscures the ground. With the immense power of a massive rocket, the entire tower begins to launch skyward in a powerful and relentless ascent, steadily defying gravity. Shockwaves of flame and debris erupt from the tower's foundation, creating a surreal and terrifying atmosphere of destruction.", + "seed": 15288, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_2/index_1145_seed_15288_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1958906540647149, + "vlm_logprob": -0.1958906540647149, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.1958906540647149, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014271103776991367 + ], + "opposite_logprob_tokens": [ + -4.389271259307861 + ], + "logprob_reward": -0.014271103776991367, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014271103776991367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14410045742988586, + 0.030103763565421104, + 0.016194293275475502, + 0.10187896341085434, + 0.18586823344230652, + 0.033502183854579926, + 0.035555530339479446, + 0.33619871735572815, + 0.03602852672338486, + 0.0805693119764328 + ], + "credit_map": [ + 0.019746463745832443, + 0.013345474377274513, + 0.009834321215748787, + 0.042861707508563995, + 0.01806984469294548, + 0.008802825585007668, + 0.024610579013824463, + 0.00682924035936594, + 0.0035099696833640337, + 0.002823724877089262, + 0.0028609398286789656, + 0.007217192091047764, + 0.004855136387050152, + 0.0022351841907948256, + 0.0051827384158968925, + 0.0014188782079145312, + 0.001496142242103815, + 0.0019358948338776827, + 0.0020656322594732046, + 0.0023532649502158165, + 0.0029687038622796535, + 0.0014842667151242495, + 0.002954361727461219, + 0.0009360249969176948, + 0.010952727869153023, + 0.007250507362186909, + 0.01519034430384636, + 0.024754809215664864, + 0.015867041423916817, + 0.007937473244965076, + 0.013365488499403, + 0.006560570560395718, + 0.018650071695446968, + 0.013726901262998581, + 0.0682474672794342, + 0.026824036613106728, + 0.02582351304590702, + 0.009047244675457478, + 0.01316213607788086, + 0.010386875830590725, + 0.0037645965348929167, + 0.0036498329136520624, + 0.004472036380320787, + 0.004230414051562548, + 0.005537703633308411, + 0.0028000271413475275, + 0.005611767526715994, + 0.0034358063712716103, + 0.003560109529644251, + 0.003546033287420869, + 0.00535739166662097, + 0.005037222057580948, + 0.004225021693855524, + 0.004949777387082577, + 0.003775667631998658, + 0.005104307550936937, + 0.06027466803789139, + 0.03944827988743782, + 0.04800921678543091, + 0.034524161368608475, + 0.032572075724601746, + 0.02725229226052761, + 0.06804993748664856, + 0.026068078354001045, + 0.004638747777789831, + 0.0035698481369763613, + 0.007281586527824402, + 0.003774323733523488, + 0.0030870670452713966, + 0.0031501196790486574, + 0.006419830489903688, + 0.004107003565877676, + 0.005952669307589531, + 0.007233328651636839, + 0.0068359654396772385, + 0.0070478240959346294, + 0.02211458794772625, + 0.014235148206353188, + 0.0069888862781226635, + 0.01016089878976345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.014271103776991367 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004118057899177074 + ], + "opposite_logprob_tokens": [ + -5.504117965698242 + ], + "logprob_reward": -0.004118057899177074, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004118057899177074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16324399411678314, + 0.043989069759845734, + 0.04495087265968323, + 0.11611641198396683, + 0.21824634075164795, + 0.06857208162546158, + 0.04289042204618454, + 0.13036350905895233, + 0.09477415680885315, + 0.07685316354036331 + ], + "credit_map": [ + 0.017858780920505524, + 0.01768677495419979, + 0.017988204956054688, + 0.04394859820604324, + 0.01994592323899269, + 0.014970789663493633, + 0.024924203753471375, + 0.005920724011957645, + 0.005508655682206154, + 0.0043142810463905334, + 0.009013497270643711, + 0.007348295301198959, + 0.005114180035889149, + 0.004015414044260979, + 0.006660511717200279, + 0.0020142316352576017, + 0.00671541690826416, + 0.003932096995413303, + 0.007932539097964764, + 0.008854675106704235, + 0.005116533022373915, + 0.003735128790140152, + 0.005956555716693401, + 0.0027079256251454353, + 0.017757317051291466, + 0.008228895254433155, + 0.008428717032074928, + 0.037464797496795654, + 0.019437937065958977, + 0.004697073716670275, + 0.01408222783356905, + 0.006019452586770058, + 0.02489808388054371, + 0.025584660470485687, + 0.048281148076057434, + 0.02404347062110901, + 0.03149613365530968, + 0.02693822979927063, + 0.01512808259576559, + 0.02187652327120304, + 0.01168654952198267, + 0.007068059407174587, + 0.009669393301010132, + 0.00799407996237278, + 0.009701094590127468, + 0.006552238017320633, + 0.007399271242320538, + 0.008501392789185047, + 0.005162608344107866, + 0.005496152676641941, + 0.005248769652098417, + 0.004366335459053516, + 0.005639420356601477, + 0.006193865556269884, + 0.004636571742594242, + 0.006146698724478483, + 0.01982968859374523, + 0.013125143945217133, + 0.03164854273200035, + 0.013457519933581352, + 0.012751908041536808, + 0.01190758217126131, + 0.015304632484912872, + 0.012338495813310146, + 0.014451028779149055, + 0.011178867891430855, + 0.018720125779509544, + 0.009115488268435001, + 0.008337019942700863, + 0.009999065659940243, + 0.011942262761294842, + 0.011030299589037895, + 0.0074117835611104965, + 0.006318957079201937, + 0.01185452938079834, + 0.011127765290439129, + 0.011708283796906471, + 0.010253218933939934, + 0.008707285858690739, + 0.009471329860389233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.004118057899177074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545713223516941 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545713223516941, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545713223516941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043367061764001846, + 0.01800018548965454, + 0.017251869663596153, + 0.1051354706287384, + 0.11354158818721771, + 0.027231745421886444, + 0.048058923333883286, + 0.14116914570331573, + 0.17137521505355835, + 0.3148687779903412 + ], + "credit_map": [ + 0.006014411337673664, + 0.008404606021940708, + 0.005922060925513506, + 0.005548904184252024, + 0.004842389840632677, + 0.005696260370314121, + 0.00480954023078084, + 0.002128888387233019, + 0.0021964204497635365, + 0.0028213453479111195, + 0.001973893027752638, + 0.0024009172338992357, + 0.0024509751237928867, + 0.001249961438588798, + 0.003640800714492798, + 0.0012658730847761035, + 0.0016900859773159027, + 0.0022329811472445726, + 0.0022807742934674025, + 0.003194049932062626, + 0.002579060848802328, + 0.0012386072194203734, + 0.003066669451072812, + 0.0009696394554339349, + 0.016748476773500443, + 0.007956611923873425, + 0.0082051707431674, + 0.032835591584444046, + 0.015336891636252403, + 0.005228906869888306, + 0.013291288167238235, + 0.005532539915293455, + 0.010714340023696423, + 0.010726716369390488, + 0.03620848059654236, + 0.016372499987483025, + 0.014519267715513706, + 0.007526999805122614, + 0.008082380518317223, + 0.00939090084284544, + 0.003094273619353771, + 0.0035104311536997557, + 0.003531121416017413, + 0.0029098670929670334, + 0.004410619381815195, + 0.0026224469766020775, + 0.0041534327901899815, + 0.002999553456902504, + 0.003755585290491581, + 0.0050704893656075, + 0.007956829853355885, + 0.005597786046564579, + 0.005448448471724987, + 0.00597799988463521, + 0.006373065989464521, + 0.007878721691668034, + 0.023051386699080467, + 0.014200293458998203, + 0.033287566155195236, + 0.01331239566206932, + 0.012555096298456192, + 0.009842002764344215, + 0.019037723541259766, + 0.015882674604654312, + 0.02224324271082878, + 0.02157159149646759, + 0.0366821326315403, + 0.017014749348163605, + 0.01374193374067545, + 0.014631427824497223, + 0.024294288828969002, + 0.021195847541093826, + 0.033099181950092316, + 0.03719302639365196, + 0.038790229707956314, + 0.0434173122048378, + 0.05113520473241806, + 0.040958505123853683, + 0.04030993580818176, + 0.029965393245220184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.012545713223516941 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742591977119446 + ], + "opposite_logprob_tokens": [ + -0.9742591977119446 + ], + "logprob_reward": -0.4742591977119446, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742591977119446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04719197377562523, + 0.04627171903848648, + 0.0334850437939167, + 0.07833218574523926, + 0.3356536030769348, + 0.062215883284807205, + 0.0743858739733696, + 0.12297357618808746, + 0.052498552948236465, + 0.14699162542819977 + ], + "credit_map": [ + 0.006980282720178366, + 0.0062513709999620914, + 0.004570845980197191, + 0.011141541413962841, + 0.005462557077407837, + 0.0042499881237745285, + 0.0063343229703605175, + 0.002201063558459282, + 0.006746862083673477, + 0.0032722095493227243, + 0.003061339957639575, + 0.014375495724380016, + 0.0056776790879666805, + 0.0023202996235340834, + 0.008869816549122334, + 0.0019480119226500392, + 0.0035396800376474857, + 0.00425700331106782, + 0.002405053237453103, + 0.006930484902113676, + 0.005887498147785664, + 0.002638918347656727, + 0.005939149297773838, + 0.00188725464977324, + 0.007617379073053598, + 0.004986950661987066, + 0.008367775939404964, + 0.02023308165371418, + 0.013783267699182034, + 0.006337280385196209, + 0.011877703480422497, + 0.005128742195665836, + 0.03208879381418228, + 0.0245542973279953, + 0.13010990619659424, + 0.05497230961918831, + 0.04671034216880798, + 0.012877514585852623, + 0.021550333127379417, + 0.012790072709321976, + 0.007099503185600042, + 0.008669504895806313, + 0.00880985613912344, + 0.006877674721181393, + 0.010086526162922382, + 0.008984541520476341, + 0.007561695761978626, + 0.0041265771724283695, + 0.00797053799033165, + 0.007778793573379517, + 0.011998574249446392, + 0.01029279362410307, + 0.010360381565988064, + 0.008444678969681263, + 0.008534026332199574, + 0.009006083011627197, + 0.017311645671725273, + 0.009503211826086044, + 0.0265654306858778, + 0.013512839563190937, + 0.010689213871955872, + 0.0071233417838811874, + 0.025233402848243713, + 0.013034479692578316, + 0.007312458474189043, + 0.006130825728178024, + 0.009982259944081306, + 0.005655842367559671, + 0.00433911569416523, + 0.004110120236873627, + 0.006831663195043802, + 0.00813626404851675, + 0.014392038807272911, + 0.01578119955956936, + 0.016383126378059387, + 0.018879076465964317, + 0.023273056373000145, + 0.02340567857027054, + 0.018242187798023224, + 0.01663525030016899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.4742591977119446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742591977119446 + ], + "opposite_logprob_tokens": [ + -0.9742591977119446 + ], + "logprob_reward": -0.4742591977119446, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742591977119446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06316319108009338, + 0.03984098136425018, + 0.023648014292120934, + 0.057905811816453934, + 0.4205974340438843, + 0.05948461592197418, + 0.07998575270175934, + 0.10730674862861633, + 0.04919365793466568, + 0.0988737940788269 + ], + "credit_map": [ + 0.008963087573647499, + 0.00940201710909605, + 0.007373213302344084, + 0.011706773191690445, + 0.006581935565918684, + 0.007208285853266716, + 0.008591977879405022, + 0.0033359029330313206, + 0.004549778997898102, + 0.0038052245508879423, + 0.003505471395328641, + 0.01166851632297039, + 0.00498353224247694, + 0.0031438511796295643, + 0.0065361205488443375, + 0.0016484917141497135, + 0.0029223477467894554, + 0.002898545004427433, + 0.004290036857128143, + 0.002848721807822585, + 0.0025492373388260603, + 0.003911070059984922, + 0.002651933580636978, + 0.0015761246904730797, + 0.00675718579441309, + 0.004106429871171713, + 0.005767620634287596, + 0.016138995066285133, + 0.009403403848409653, + 0.003236420452594757, + 0.007772416342049837, + 0.004723344463855028, + 0.03938768059015274, + 0.03125018626451492, + 0.15751433372497559, + 0.06184497848153114, + 0.0572153776884079, + 0.02445073425769806, + 0.028174247592687607, + 0.02075990103185177, + 0.006382606457918882, + 0.007718932349234819, + 0.007690045051276684, + 0.006228337530046701, + 0.008814402855932713, + 0.00681174173951149, + 0.010106312111020088, + 0.005732241086661816, + 0.0065279873088002205, + 0.01039455458521843, + 0.009723138064146042, + 0.009851873852312565, + 0.011105874553322792, + 0.009817471727728844, + 0.009962495416402817, + 0.0126023655757308, + 0.022151583805680275, + 0.013193151913583279, + 0.017457490786910057, + 0.009388046339154243, + 0.010857265442609787, + 0.008654126897454262, + 0.016255736351013184, + 0.009349353611469269, + 0.010538145899772644, + 0.005200895480811596, + 0.007304247468709946, + 0.004236552864313126, + 0.003815284464508295, + 0.0048614912666380405, + 0.005604562349617481, + 0.007632480002939701, + 0.007668569218367338, + 0.010652847588062286, + 0.011554933153092861, + 0.013563684187829494, + 0.016713956370949745, + 0.015404271893203259, + 0.012861165218055248, + 0.010454367846250534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.4742591977119446 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle, wide-angle shot captures the scene, the camera initially static to establish the monumental scale, but then slowly, deliberately tilting upward to follow the unfolding action. At dusk, beneath a turbulent sky of deep purple and scarred orange clouds, Paris's iconic iron lattice tower stands tall, its intricate cast-iron latticework rendered in sharp, lifelike detail. Suddenly, a violent and profound tremor sweeps through the vast structure, causing dust and small metal fragments to break free from its beams. The tower's base is instantly engulfed in a roaring inferno of brilliant orange and yellow flames, releasing a vast, billowing cloud of thick, black and gray smoke that obscures the ground. With the immense power of a massive rocket, the entire tower begins to launch skyward in a powerful and relentless ascent, steadily defying gravity. Shockwaves of flame and debris erupt from the tower's foundation, creating a surreal and terrifying atmosphere of destruction.", + "seed": 15289, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_2/index_1145_seed_15289_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13034565318375826, + "vlm_logprob": -0.13034565318375826, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.13034565318375826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022792508825659752 + ], + "opposite_logprob_tokens": [ + -3.8977925777435303 + ], + "logprob_reward": -0.022792508825659752, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022792508825659752, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12138005346059799, + 0.02200618013739586, + 0.04290904477238655, + 0.11679882556200027, + 0.10527057945728302, + 0.08010164648294449, + 0.18798187375068665, + 0.19627167284488678, + 0.0869661346077919, + 0.040313925594091415 + ], + "credit_map": [ + 0.02161930315196514, + 0.011227171868085861, + 0.00859537348151207, + 0.005836596246808767, + 0.013373044319450855, + 0.013065868988633156, + 0.006339501589536667, + 0.041323207318782806, + 0.0022702794522047043, + 0.0037423784378916025, + 0.002341657876968384, + 0.0023039511870592833, + 0.004542483948171139, + 0.0032080821692943573, + 0.0017816724721342325, + 0.0018156770383939147, + 0.008353054523468018, + 0.0029856753535568714, + 0.0038181478157639503, + 0.0029733912087976933, + 0.013158777728676796, + 0.005525872111320496, + 0.00349594047293067, + 0.0025981892831623554, + 0.039495088160037994, + 0.004782767500728369, + 0.03412223607301712, + 0.004507773090153933, + 0.005389658734202385, + 0.009542168118059635, + 0.004404708743095398, + 0.014554440975189209, + 0.00822095200419426, + 0.011748116463422775, + 0.011531719006597996, + 0.010598938912153244, + 0.015066585503518581, + 0.015983780845999718, + 0.02227688767015934, + 0.009843611158430576, + 0.005373516585677862, + 0.012801218777894974, + 0.008961683139204979, + 0.011966033838689327, + 0.010621225461363792, + 0.006681196391582489, + 0.01161279808729887, + 0.012083979323506355, + 0.028556741774082184, + 0.01867559179663658, + 0.027642011642456055, + 0.02144762873649597, + 0.021527986973524094, + 0.017705585807561874, + 0.03064790554344654, + 0.021778441965579987, + 0.021054141223430634, + 0.030699409544467926, + 0.02215017005801201, + 0.03026108257472515, + 0.02579878829419613, + 0.0204070545732975, + 0.026079075410962105, + 0.019821971654891968, + 0.00815494079142809, + 0.010123252868652344, + 0.012530675157904625, + 0.010940996930003166, + 0.01256746519356966, + 0.014012840576469898, + 0.010428315959870815, + 0.00820765271782875, + 0.006797911133617163, + 0.005071818828582764, + 0.0019427647348493338, + 0.004440941847860813, + 0.014168592169880867, + 0.0026289490051567554, + 0.0019099776400253177, + 0.003352972213178873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.022792508825659752 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373095020651817 + ], + "opposite_logprob_tokens": [ + -3.408730983734131 + ], + "logprob_reward": -0.03373095020651817, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373095020651817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13520601391792297, + 0.05940786749124527, + 0.10715381801128387, + 0.1255064159631729, + 0.06775923073291779, + 0.05883904546499252, + 0.11687644571065903, + 0.24631421267986298, + 0.05231538787484169, + 0.030621567741036415 + ], + "credit_map": [ + 0.024778561666607857, + 0.012762289494276047, + 0.008746668696403503, + 0.00701970886439085, + 0.014664019457995892, + 0.01509244553744793, + 0.006607978139072657, + 0.045534346252679825, + 0.0053611453622579575, + 0.008679703809320927, + 0.003420531749725342, + 0.007623692974448204, + 0.010943105444312096, + 0.011705376207828522, + 0.006359327584505081, + 0.005314984824508429, + 0.0208414476364851, + 0.006648709066212177, + 0.007444277871400118, + 0.006672750227153301, + 0.03696797415614128, + 0.015739820897579193, + 0.007558346726000309, + 0.005280502140522003, + 0.04042103886604309, + 0.005559345241636038, + 0.03208017349243164, + 0.005069749895483255, + 0.006418643984943628, + 0.011611774563789368, + 0.00487451022490859, + 0.0194711834192276, + 0.011425028555095196, + 0.009690089151263237, + 0.007702221162617207, + 0.005346665624529123, + 0.007186682429164648, + 0.013729933649301529, + 0.005308785475790501, + 0.007369820959866047, + 0.0058433860540390015, + 0.00848525669425726, + 0.007059978321194649, + 0.008440920151770115, + 0.006526388693600893, + 0.007312034722417593, + 0.007109377533197403, + 0.00806170329451561, + 0.015459333546459675, + 0.01188689935952425, + 0.017740242183208466, + 0.014202856458723545, + 0.013087672181427479, + 0.010696941055357456, + 0.019062289968132973, + 0.014740217477083206, + 0.026934554800391197, + 0.04114832729101181, + 0.025189699605107307, + 0.0368628203868866, + 0.034012023359537125, + 0.02489120326936245, + 0.03504668548703194, + 0.02222888171672821, + 0.0052847168408334255, + 0.006723353173583746, + 0.007239275611937046, + 0.0069195362739264965, + 0.007564011961221695, + 0.0071531482972204685, + 0.006319047417491674, + 0.0051123034209012985, + 0.007327036466449499, + 0.0034932447597384453, + 0.002345908433198929, + 0.0030713200103491545, + 0.008245068602263927, + 0.0022424522321671247, + 0.0017140581039711833, + 0.0021824778523296118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.03373095020651817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595568772405386 + ], + "opposite_logprob_tokens": [ + -5.130955696105957 + ], + "logprob_reward": -0.00595568772405386, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595568772405386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22309564054012299, + 0.05468195304274559, + 0.02645600028336048, + 0.17824961245059967, + 0.040796782821416855, + 0.03267445042729378, + 0.07533840835094452, + 0.1373591125011444, + 0.11103102564811707, + 0.1203169971704483 + ], + "credit_map": [ + 0.04261816293001175, + 0.02059127762913704, + 0.013335438445210457, + 0.009249215014278889, + 0.0234957505017519, + 0.023071393370628357, + 0.008775459602475166, + 0.08195894956588745, + 0.005565420258790255, + 0.009987127967178822, + 0.004483118187636137, + 0.005474605597555637, + 0.01138230413198471, + 0.008740311488509178, + 0.004656758159399033, + 0.004392312373965979, + 0.003731335513293743, + 0.0028299111872911453, + 0.005927411839365959, + 0.002205884549766779, + 0.003469928866252303, + 0.003058473113924265, + 0.0020275688730180264, + 0.0032054856419563293, + 0.06414737552404404, + 0.006950586102902889, + 0.06683212518692017, + 0.008000372909009457, + 0.007003820966929197, + 0.007988468743860722, + 0.006306374911218882, + 0.01102051418274641, + 0.004609784577041864, + 0.0041939555667340755, + 0.003696802770718932, + 0.002781814197078347, + 0.00565529428422451, + 0.012200342491269112, + 0.0037911145482212305, + 0.0038676713593304157, + 0.0030983074102550745, + 0.003537919372320175, + 0.004571027588099241, + 0.0031248731538653374, + 0.005862733349204063, + 0.0032278867438435555, + 0.004846638534218073, + 0.004405065439641476, + 0.009151933714747429, + 0.007610877510160208, + 0.009436269290745258, + 0.011995915323495865, + 0.007203544490039349, + 0.006969263311475515, + 0.011990693397819996, + 0.010979917831718922, + 0.015936287119984627, + 0.02587968111038208, + 0.014917737804353237, + 0.019107675179839134, + 0.017864976078271866, + 0.014044675976037979, + 0.01800217479467392, + 0.011605911888182163, + 0.014040688052773476, + 0.019384663552045822, + 0.014681047759950161, + 0.01028284803032875, + 0.011853234842419624, + 0.01960638351738453, + 0.010603182017803192, + 0.01057897787541151, + 0.023886237293481827, + 0.01671982929110527, + 0.011255557648837566, + 0.017121626064181328, + 0.022078434005379677, + 0.008082034066319466, + 0.00975584052503109, + 0.011417440138757229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.00595568772405386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288915693759918 + ], + "opposite_logprob_tokens": [ + -1.0538915395736694 + ], + "logprob_reward": -0.4288915693759918, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288915693759918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1417774111032486, + 0.04216044768691063, + 0.04215048626065254, + 0.075358085334301, + 0.09019827842712402, + 0.06913363933563232, + 0.11615482717752457, + 0.1894703507423401, + 0.10556163638830185, + 0.128034770488739 + ], + "credit_map": [ + 0.027610337361693382, + 0.01245102658867836, + 0.008839376270771027, + 0.006404960993677378, + 0.014734648168087006, + 0.014718513935804367, + 0.006512179970741272, + 0.05050637573003769, + 0.005817623343318701, + 0.004758819472044706, + 0.004080579616129398, + 0.004907634109258652, + 0.006474539637565613, + 0.00872640497982502, + 0.004001593217253685, + 0.0033932591322809458, + 0.007710300851613283, + 0.0029316400177776814, + 0.004314629826694727, + 0.0036181313917040825, + 0.010675453580915928, + 0.006205130834132433, + 0.0036964528262615204, + 0.0029987492598593235, + 0.021641388535499573, + 0.004074083641171455, + 0.023812556639313698, + 0.005380375310778618, + 0.005562711972743273, + 0.004594721831381321, + 0.00431723240762949, + 0.005975024309009314, + 0.008038888685405254, + 0.008126894012093544, + 0.01033091451972723, + 0.006605764850974083, + 0.012896797619760036, + 0.022064508870244026, + 0.012785607948899269, + 0.009348909370601177, + 0.005055628251284361, + 0.011595139279961586, + 0.008277631364762783, + 0.009980450384318829, + 0.00837391521781683, + 0.0058201393112540245, + 0.00937928631901741, + 0.010651453398168087, + 0.01071502361446619, + 0.01098365243524313, + 0.01663900353014469, + 0.01814696006476879, + 0.011911711655557156, + 0.009196960367262363, + 0.01986502669751644, + 0.018696501851081848, + 0.020198417827486992, + 0.030028315261006355, + 0.022694427520036697, + 0.0289335735142231, + 0.024735335260629654, + 0.019515395164489746, + 0.024519972503185272, + 0.01884492300450802, + 0.010669478215277195, + 0.0135036651045084, + 0.014674381352961063, + 0.010700005106627941, + 0.01233063917607069, + 0.021052183583378792, + 0.011221818625926971, + 0.011409473605453968, + 0.02934505231678486, + 0.01317992340773344, + 0.00552819250151515, + 0.013128044083714485, + 0.04250231757760048, + 0.008889390155673027, + 0.005633617285639048, + 0.009828243404626846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.4288915693759918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603575497865677 + ], + "opposite_logprob_tokens": [ + -1.9103575944900513 + ], + "logprob_reward": -0.1603575497865677, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603575497865677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30177682638168335, + 0.06110347434878349, + 0.21303331851959229, + 0.10672836750745773, + 0.04942856356501579, + 0.04321431368589401, + 0.06698427349328995, + 0.08737998455762863, + 0.03234584629535675, + 0.03800510987639427 + ], + "credit_map": [ + 0.05866469815373421, + 0.027577975764870644, + 0.019176064059138298, + 0.01497658435255289, + 0.0306926891207695, + 0.03234590217471123, + 0.014381919987499714, + 0.10396093875169754, + 0.007099449634552002, + 0.00598506024107337, + 0.00478331046178937, + 0.009778205305337906, + 0.008645066060125828, + 0.01341588981449604, + 0.006490906234830618, + 0.004905576817691326, + 0.04001342132687569, + 0.015005729161202908, + 0.02065589278936386, + 0.013349359855055809, + 0.06491085141897202, + 0.03213045746088028, + 0.01455540582537651, + 0.012412160634994507, + 0.027449969202280045, + 0.00412777392193675, + 0.034902676939964294, + 0.004912874195724726, + 0.006296541076153517, + 0.007630353793501854, + 0.0042573632672429085, + 0.017150798812508583, + 0.005594621878117323, + 0.0055821179412305355, + 0.006124534644186497, + 0.0036513155791908503, + 0.006908009760081768, + 0.00863317959010601, + 0.007880284450948238, + 0.005054488778114319, + 0.003701985813677311, + 0.007691571488976479, + 0.004806041717529297, + 0.006457033101469278, + 0.005429580342024565, + 0.0032199076376855373, + 0.005868243053555489, + 0.006039946805685759, + 0.005198197904974222, + 0.005504739470779896, + 0.009857122786343098, + 0.009986627846956253, + 0.008400109596550465, + 0.005453812889754772, + 0.012229294516146183, + 0.010354364290833473, + 0.009255371987819672, + 0.014368311502039433, + 0.009686501696705818, + 0.01284516230225563, + 0.011902575381100178, + 0.008637655526399612, + 0.012506186962127686, + 0.008178209885954857, + 0.00313142454251647, + 0.0035996914375573397, + 0.0041543287225067616, + 0.0038339635357260704, + 0.004532451741397381, + 0.005366360303014517, + 0.0041139679960906506, + 0.003613651730120182, + 0.00850139744579792, + 0.00484774075448513, + 0.003365669632330537, + 0.004143464379012585, + 0.00913497805595398, + 0.002381729893386364, + 0.0025797574780881405, + 0.003050369443371892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.1603575497865677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle, wide-angle shot captures the scene, the camera initially static to establish the monumental scale, but then slowly, deliberately tilting upward to follow the unfolding action. At dusk, beneath a turbulent sky of deep purple and scarred orange clouds, Paris's iconic iron lattice tower stands tall, its intricate cast-iron latticework rendered in sharp, lifelike detail. Suddenly, a violent and profound tremor sweeps through the vast structure, causing dust and small metal fragments to break free from its beams. The tower's base is instantly engulfed in a roaring inferno of brilliant orange and yellow flames, releasing a vast, billowing cloud of thick, black and gray smoke that obscures the ground. With the immense power of a massive rocket, the entire tower begins to launch skyward in a powerful and relentless ascent, steadily defying gravity. Shockwaves of flame and debris erupt from the tower's foundation, creating a surreal and terrifying atmosphere of destruction.", + "seed": 15290, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_2/index_1145_seed_15290_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.22416078061796724, + "vlm_logprob": -0.22416078061796724, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.22416078061796724, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011324767954647541 + ], + "opposite_logprob_tokens": [ + -4.636324882507324 + ], + "logprob_reward": -0.011324767954647541, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011324767954647541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10127756744623184, + 0.026341112330555916, + 0.03618653118610382, + 0.07540404796600342, + 0.42550423741340637, + 0.1049899235367775, + 0.10674354434013367, + 0.04162644222378731, + 0.036410197615623474, + 0.04551639407873154 + ], + "credit_map": [ + 0.02043227106332779, + 0.009622587822377682, + 0.005682205315679312, + 0.013869836926460266, + 0.01222744956612587, + 0.007965322583913803, + 0.008464000187814236, + 0.023013895377516747, + 0.004835240077227354, + 0.0028779946733266115, + 0.0022298323456197977, + 0.003712550736963749, + 0.002370918169617653, + 0.002768728183582425, + 0.0018835000228136778, + 0.005662344861775637, + 0.00584867550060153, + 0.004775789100676775, + 0.002081839833408594, + 0.0048489049077034, + 0.004257720895111561, + 0.0035877935588359833, + 0.0032094870693981647, + 0.007576317526400089, + 0.01782512106001377, + 0.0059948996640741825, + 0.0026108890306204557, + 0.004350763279944658, + 0.0064589823596179485, + 0.028832728043198586, + 0.0041962903924286366, + 0.005134365521371365, + 0.04375641420483589, + 0.1041116863489151, + 0.03579292818903923, + 0.05329958721995354, + 0.04707803577184677, + 0.061653267592191696, + 0.04753517732024193, + 0.032277122139930725, + 0.009618520736694336, + 0.016407901421189308, + 0.014247613027691841, + 0.01218930073082447, + 0.01429422851651907, + 0.013423620723187923, + 0.014993960969150066, + 0.009814771823585033, + 0.01042355876415968, + 0.023828372359275818, + 0.018968934193253517, + 0.010822523385286331, + 0.012874468229711056, + 0.006444687955081463, + 0.01280833687633276, + 0.010572655126452446, + 0.0036333673633635044, + 0.004698907025158405, + 0.008371572941541672, + 0.004407859407365322, + 0.006239603273570538, + 0.006425404455512762, + 0.0036837372463196516, + 0.004165991209447384, + 0.0028589563444256783, + 0.00409937696531415, + 0.0028982225339859724, + 0.005654314067214727, + 0.0038926145061850548, + 0.004531588405370712, + 0.007367310579866171, + 0.005107810255140066, + 0.004972794558852911, + 0.005757632199674845, + 0.006448870059102774, + 0.006802212446928024, + 0.0035358492750674486, + 0.005342213436961174, + 0.007488226518034935, + 0.005168589763343334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.011324767954647541 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006774434354156256 + ], + "opposite_logprob_tokens": [ + -5.006774425506592 + ], + "logprob_reward": -0.006774434354156256, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006774434354156256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14112384617328644, + 0.06871595233678818, + 0.046243149787187576, + 0.06607282906770706, + 0.17245718836784363, + 0.0676288977265358, + 0.21042896807193756, + 0.12286604195833206, + 0.04972629249095917, + 0.05473683029413223 + ], + "credit_map": [ + 0.01488533616065979, + 0.008865805342793465, + 0.006604120600968599, + 0.007674876134842634, + 0.04994431883096695, + 0.014188975095748901, + 0.01064487174153328, + 0.028315545991063118, + 0.014281852170825005, + 0.0063601029105484486, + 0.006244876887649298, + 0.00455170962959528, + 0.005343188066035509, + 0.011173190549015999, + 0.0032752221450209618, + 0.017485810443758965, + 0.008905261754989624, + 0.004761779215186834, + 0.006414793897420168, + 0.004622364416718483, + 0.004209175240248442, + 0.00428353063762188, + 0.004982565995305777, + 0.008063679561018944, + 0.013654187321662903, + 0.007183850277215242, + 0.005095220636576414, + 0.004792286083102226, + 0.006130596622824669, + 0.015561823733150959, + 0.00510161928832531, + 0.008553247898817062, + 0.028812451288104057, + 0.031553152948617935, + 0.0157921202480793, + 0.022935058921575546, + 0.01563457027077675, + 0.021983733400702477, + 0.018209269270300865, + 0.017536835744976997, + 0.005357983987778425, + 0.011883175000548363, + 0.009069131687283516, + 0.007127909455448389, + 0.010032907128334045, + 0.008047951385378838, + 0.009279115125536919, + 0.0068307192996144295, + 0.027828887104988098, + 0.03354779630899429, + 0.04175672307610512, + 0.02445741556584835, + 0.02723456360399723, + 0.017478419467806816, + 0.02041827328503132, + 0.017706885933876038, + 0.013602979481220245, + 0.01636498048901558, + 0.01763932779431343, + 0.013571924529969692, + 0.01714150980114937, + 0.016046855598688126, + 0.013213272206485271, + 0.015285196714103222, + 0.0039291069842875, + 0.004474662244319916, + 0.005047094542533159, + 0.006754618138074875, + 0.0051704938523471355, + 0.0050798943266272545, + 0.009089951403439045, + 0.010180473327636719, + 0.007005384191870689, + 0.00733424024656415, + 0.008703493513166904, + 0.005013234913349152, + 0.0055460440926253796, + 0.006625098176300526, + 0.007484392262995243, + 0.007024944759905338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.006774434354156256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032011240255087614 + ], + "opposite_logprob_tokens": [ + -5.753201007843018 + ], + "logprob_reward": -0.0032011240255087614, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032011240255087614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048492711037397385, + 0.029343299567699432, + 0.035617247223854065, + 0.014786251820623875, + 0.11732755601406097, + 0.06042926758527756, + 0.2825755774974823, + 0.15183871984481812, + 0.11684753745794296, + 0.14274181425571442 + ], + "credit_map": [ + 0.0082576684653759, + 0.0032758351881057024, + 0.0031440667808055878, + 0.002862335415557027, + 0.009103373624384403, + 0.00727312033995986, + 0.002907010493800044, + 0.011669301427900791, + 0.0037631234154105186, + 0.004852998070418835, + 0.002641082275658846, + 0.005037335213273764, + 0.0031585742253810167, + 0.003795420750975609, + 0.002016183454543352, + 0.004078584257513285, + 0.0053419494070112705, + 0.009194079786539078, + 0.0026457554195076227, + 0.003684739349409938, + 0.004894217010587454, + 0.004171611741185188, + 0.0026612719520926476, + 0.003023622091859579, + 0.0021567477378994226, + 0.0013198880478739738, + 0.0011695652501657605, + 0.0015683288220316172, + 0.002624157350510359, + 0.00269734812900424, + 0.0016699960688129067, + 0.001580220297910273, + 0.01243259571492672, + 0.028802718967199326, + 0.010254822671413422, + 0.01508395466953516, + 0.012357305735349655, + 0.016838282346725464, + 0.01245498564094305, + 0.009102887474000454, + 0.006218837574124336, + 0.010209747590124607, + 0.008575406856834888, + 0.005982846487313509, + 0.007188360206782818, + 0.006631954573094845, + 0.008274799212813377, + 0.007347311358898878, + 0.03391081839799881, + 0.07588209956884384, + 0.07013007253408432, + 0.02438417449593544, + 0.01929388754069805, + 0.018337085843086243, + 0.018602419644594193, + 0.02203504554927349, + 0.011220081709325314, + 0.017467912286520004, + 0.019807137548923492, + 0.02074461430311203, + 0.01749218814074993, + 0.01821797899901867, + 0.01960904709994793, + 0.027279753237962723, + 0.010714211501181126, + 0.013770809397101402, + 0.012899664230644703, + 0.01559685543179512, + 0.013359492644667625, + 0.016367660835385323, + 0.021072115749120712, + 0.01306673139333725, + 0.01748371310532093, + 0.014665183611214161, + 0.024398215115070343, + 0.0166451558470726, + 0.012874973006546497, + 0.01962483674287796, + 0.016812926158308983, + 0.02023680880665779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0032011240255087614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233515501022339 + ], + "opposite_logprob_tokens": [ + -0.8983515501022339 + ], + "logprob_reward": -0.5233515501022339, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233515501022339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09261367470026016, + 0.04258399084210396, + 0.04860888794064522, + 0.03938072547316551, + 0.25532153248786926, + 0.07226959615945816, + 0.07880640774965286, + 0.19614377617835999, + 0.10281640291213989, + 0.0714549720287323 + ], + "credit_map": [ + 0.014668679796159267, + 0.008511494845151901, + 0.005013239104300737, + 0.009252671152353287, + 0.017469270154833794, + 0.008271309547126293, + 0.005725510884076357, + 0.023701496422290802, + 0.007431953214108944, + 0.004216755740344524, + 0.0028674635104835033, + 0.007366226986050606, + 0.0043477085418999195, + 0.003964010160416365, + 0.002663676394149661, + 0.009726199321448803, + 0.010538000613451004, + 0.0038095314521342516, + 0.002565772505477071, + 0.005406389012932777, + 0.004795988090336323, + 0.0034568714909255505, + 0.005187827628105879, + 0.01284850761294365, + 0.00771962059661746, + 0.0036603824701160192, + 0.002226301236078143, + 0.005253484472632408, + 0.005125381983816624, + 0.008828983642160892, + 0.003977028653025627, + 0.0025895407889038324, + 0.02762533910572529, + 0.0605388879776001, + 0.01791660487651825, + 0.024886105209589005, + 0.02731730416417122, + 0.05608317255973816, + 0.01943475566804409, + 0.021519377827644348, + 0.00741147343069315, + 0.012846184894442558, + 0.008608593605458736, + 0.007815190590918064, + 0.009910611435770988, + 0.008296502754092216, + 0.010405727662146091, + 0.00697531271725893, + 0.009284229017794132, + 0.01583344303071499, + 0.008885147050023079, + 0.0067721176892519, + 0.011516958475112915, + 0.006704932544380426, + 0.011773789301514626, + 0.008035792037844658, + 0.016970910131931305, + 0.027596382424235344, + 0.028850534930825233, + 0.0214141346514225, + 0.031818076968193054, + 0.023585474118590355, + 0.02256879024207592, + 0.02333947643637657, + 0.010651328600943089, + 0.02003725804388523, + 0.014234587550163269, + 0.010635734535753727, + 0.01130541693419218, + 0.010206450708210468, + 0.013691158965229988, + 0.012054466642439365, + 0.009064434096217155, + 0.008842218667268753, + 0.009904817678034306, + 0.00884216371923685, + 0.006098598707467318, + 0.011360514909029007, + 0.008071682415902615, + 0.009270545095205307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.5233515501022339 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the Eiffel Tower visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky colored with deep purple and scarred orange clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the tower's base burn with orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the entire tower launch skyward in an ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start static and then slowly tilt upward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761520266532898 + ], + "opposite_logprob_tokens": [ + -0.8261520266532898 + ], + "logprob_reward": -0.5761520266532898, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761520266532898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06211031228303909, + 0.032996002584695816, + 0.06992391496896744, + 0.04111624136567116, + 0.1591586023569107, + 0.10865899920463562, + 0.31445756554603577, + 0.06719473749399185, + 0.07538474351167679, + 0.06899891793727875 + ], + "credit_map": [ + 0.007336035370826721, + 0.00618927413597703, + 0.005461110733449459, + 0.005893315188586712, + 0.01173796784132719, + 0.007329941261559725, + 0.004686500411480665, + 0.013476165011525154, + 0.004319428000599146, + 0.003752527991309762, + 0.00493625458329916, + 0.004466115962713957, + 0.003373562591150403, + 0.004134845454245806, + 0.002975449664518237, + 0.005037817172706127, + 0.0074058580212295055, + 0.014061588793992996, + 0.0053874109871685505, + 0.01060203555971384, + 0.01046732161194086, + 0.007555932272225618, + 0.0066353557631373405, + 0.007808404974639416, + 0.008176964707672596, + 0.0035679645370692015, + 0.002125530270859599, + 0.0030287227127701044, + 0.003774886252358556, + 0.013378729112446308, + 0.0030235012527555227, + 0.004039940424263477, + 0.018037894740700722, + 0.024592693895101547, + 0.01089318934828043, + 0.014275317080318928, + 0.025658292695879936, + 0.03391273319721222, + 0.01689242757856846, + 0.014896050095558167, + 0.01289826724678278, + 0.016504332423210144, + 0.01607816480100155, + 0.011361801996827126, + 0.010981939733028412, + 0.015328537672758102, + 0.01403260137885809, + 0.011473353952169418, + 0.045039158314466476, + 0.07267601788043976, + 0.0805361270904541, + 0.031320538371801376, + 0.02483196370303631, + 0.016690270975232124, + 0.020066125318408012, + 0.023297350853681564, + 0.00768935214728117, + 0.009342360310256481, + 0.0125351557508111, + 0.007957935333251953, + 0.009973650798201561, + 0.00707145594060421, + 0.006266172043979168, + 0.006358649116009474, + 0.006353577598929405, + 0.011187833733856678, + 0.008736523799598217, + 0.008950413204729557, + 0.008636713027954102, + 0.009539954364299774, + 0.013637430034577847, + 0.00834229588508606, + 0.00631918478757143, + 0.007233535870909691, + 0.007878080010414124, + 0.010218039155006409, + 0.005022549070417881, + 0.01017662975937128, + 0.013599874451756477, + 0.008551022037863731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.5761520266532898 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_3/index_tensor([2264, 2264])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_3/index_tensor([2264, 2264])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d3fa630103c43d51b8f347734a7be23272c1c13a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_3/index_tensor([2264, 2264])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fluid, eye-level tracking shot follows a majestic hybrid creature as it gracefully soars through a dramatic sunset sky. The creature possesses the powerful, muscular body of a Tyrannosaurus Rex, covered in fine, interlocking scales of deep emerald green with a bronze sheen. Its mighty legs hug its body, and a long, heavy tail flows behind it, delicately adjusting its balance. The pterosaur's massive, leathery wings blend seamlessly into its powerful shoulders, their membranes a rich, earthy brown, revealing a delicate network of veins as the light filters through them. The creature glides with effortless grace, wings spread wide and steady, sailing through fluffy clouds soaked with the day's last rays. Their undersides glow a fiery orange, their crowns brushed with soft pinks and lavenders. A low-angled golden sun acts as a brilliant key light, capturing the sharp edges of each individual scale and creating a glowing halo around the creature's entire silhouette, imbuing the scene with a breathtaking, pristine tranquility.", + "seed": 18581, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_3/index_2264_seed_18581_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.057938839122653005, + "vlm_logprob": -0.057938839122653005, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.057938839122653005, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034393906593322754 + ], + "opposite_logprob_tokens": [ + -3.534393787384033 + ], + "logprob_reward": -0.034393906593322754, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.034393906593322754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06246578320860863, + 0.20397496223449707, + 0.05428262799978256, + 0.06606880575418472, + 0.08463985472917557, + 0.036397311836481094, + 0.09730007499456406, + 0.0248731542378664, + 0.021077480167150497, + 0.34891995787620544 + ], + "credit_map": [ + 0.007967936806380749, + 0.008284889161586761, + 0.008043651469051838, + 0.005764719098806381, + 0.008795536123216152, + 0.00745120644569397, + 0.009306864812970161, + 0.006850976962596178, + 0.03306305408477783, + 0.013719498179852962, + 0.030078331008553505, + 0.020177341997623444, + 0.018590712919831276, + 0.029217280447483063, + 0.026019394397735596, + 0.033109355717897415, + 0.00588464317843318, + 0.005316826049238443, + 0.008022075518965721, + 0.007504811976104975, + 0.008007795549929142, + 0.0059967124834656715, + 0.005891985259950161, + 0.007657780312001705, + 0.012545057572424412, + 0.0067206076346337795, + 0.009384458884596825, + 0.01086552906781435, + 0.0055009848438203335, + 0.00832774955779314, + 0.005485909059643745, + 0.0072385044768452644, + 0.009889786131680012, + 0.006382596213370562, + 0.009232277050614357, + 0.01787075400352478, + 0.005512121133506298, + 0.0047386204823851585, + 0.005693483166396618, + 0.02532021515071392, + 0.004986034706234932, + 0.005343105643987656, + 0.005098823457956314, + 0.008795959874987602, + 0.0023255564738065004, + 0.002335736295208335, + 0.003719717264175415, + 0.003792375326156616, + 0.011492564342916012, + 0.014009996317327023, + 0.00809340551495552, + 0.013707288540899754, + 0.01017718855291605, + 0.00962531566619873, + 0.01324364822357893, + 0.016950665041804314, + 0.00374027737416327, + 0.0034553762525320053, + 0.002280783373862505, + 0.002239679219201207, + 0.002976704388856888, + 0.003492251969873905, + 0.0037913559935986996, + 0.0028967270627617836, + 0.003077386412769556, + 0.0024126586504280567, + 0.0016883824719116092, + 0.003655749838799238, + 0.0021578145679086447, + 0.0020159974228590727, + 0.0035632336512207985, + 0.0025062572676688433, + 0.035115666687488556, + 0.04094419628381729, + 0.04997007176280022, + 0.033875275403261185, + 0.0345904640853405, + 0.06294674426317215, + 0.041845209896564484, + 0.049632322043180466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.034393906593322754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005482912063599 + ], + "opposite_logprob_tokens": [ + -2.695054769515991 + ], + "logprob_reward": -0.07005482912063599, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005482912063599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06668699532747269, + 0.06940861791372299, + 0.0710301548242569, + 0.0979963019490242, + 0.054141078144311905, + 0.03736433759331703, + 0.04526358097791672, + 0.03196558728814125, + 0.0454639196395874, + 0.4806794822216034 + ], + "credit_map": [ + 0.00505517004057765, + 0.00865099299699068, + 0.009595179930329323, + 0.006035707425326109, + 0.0067763072438538074, + 0.015263035893440247, + 0.008723406121134758, + 0.0065871975384652615, + 0.004788906313478947, + 0.01108180359005928, + 0.009171838872134686, + 0.009382171556353569, + 0.008543798699975014, + 0.00997505709528923, + 0.006693248171359301, + 0.009771795943379402, + 0.008420269005000591, + 0.006455409340560436, + 0.009810556657612324, + 0.0076576718129217625, + 0.00784366950392723, + 0.00850796326994896, + 0.008681531995534897, + 0.013653086498379707, + 0.012876756489276886, + 0.006745422724634409, + 0.0208175890147686, + 0.012732858769595623, + 0.007509865798056126, + 0.012109421193599701, + 0.008044401183724403, + 0.01715998351573944, + 0.006065045949071646, + 0.003876294707879424, + 0.00882924534380436, + 0.013233165256679058, + 0.002765786135569215, + 0.0037538164760917425, + 0.004427401348948479, + 0.011190321296453476, + 0.004218448884785175, + 0.0038795596919953823, + 0.005520694889128208, + 0.008860657922923565, + 0.0023663088213652372, + 0.0034812865778803825, + 0.0040870350785553455, + 0.00495034409686923, + 0.006907119881361723, + 0.006035295780748129, + 0.00415499322116375, + 0.007075528148561716, + 0.005298083182424307, + 0.004495411179959774, + 0.004044048488140106, + 0.007253098767250776, + 0.004827720113098621, + 0.004240845330059528, + 0.002863408299162984, + 0.002678132150322199, + 0.004077348858118057, + 0.0058831884525716305, + 0.004639199003577232, + 0.0027557439170777798, + 0.006107005290687084, + 0.00502484617754817, + 0.0037073465064167976, + 0.005356431473046541, + 0.006644063163548708, + 0.005362444091588259, + 0.007984722964465618, + 0.005277057643979788, + 0.05069689080119133, + 0.04455862194299698, + 0.06361176818609238, + 0.04707687720656395, + 0.05684235692024231, + 0.06442084908485413, + 0.07120320945978165, + 0.08226887136697769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.07005482912063599 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743056699633598 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006743056699633598, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743056699633598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07743562012910843, + 0.15172499418258667, + 0.06933122873306274, + 0.1325843781232834, + 0.042531561106443405, + 0.0763113871216774, + 0.14283668994903564, + 0.041115902364254, + 0.03225830942392349, + 0.23386994004249573 + ], + "credit_map": [ + 0.005937600042670965, + 0.0070120724849402905, + 0.012552537955343723, + 0.007466241717338562, + 0.007062963675707579, + 0.015332290902733803, + 0.01381262019276619, + 0.008259291760623455, + 0.019324641674757004, + 0.019384047016501427, + 0.019324954599142075, + 0.018727147951722145, + 0.012953240424394608, + 0.018680734559893608, + 0.02065778523683548, + 0.022672437131404877, + 0.006532777566462755, + 0.004865485243499279, + 0.011244365014135838, + 0.012715185061097145, + 0.007674845866858959, + 0.007228408008813858, + 0.009083902463316917, + 0.00998626183718443, + 0.025930577889084816, + 0.013932608999311924, + 0.018376916646957397, + 0.01801609992980957, + 0.012812086381018162, + 0.015631260350346565, + 0.012177472934126854, + 0.01570734940469265, + 0.00432216702029109, + 0.0026925995480269194, + 0.008430799469351768, + 0.008213897235691547, + 0.003333353903144598, + 0.004456191323697567, + 0.0042998348362743855, + 0.006782719399780035, + 0.00836650375276804, + 0.00789634045213461, + 0.012009312398731709, + 0.017136460170149803, + 0.006245928350836039, + 0.007078635040670633, + 0.00786740891635418, + 0.009710799902677536, + 0.015488036908209324, + 0.02007478103041649, + 0.01677139662206173, + 0.02451326511800289, + 0.01623023860156536, + 0.015624191612005234, + 0.0131229842081666, + 0.021011797711253166, + 0.007330543361604214, + 0.005879370030015707, + 0.005200179759413004, + 0.005101494025439024, + 0.0030419982504099607, + 0.0049087232910096645, + 0.005890048574656248, + 0.0037635474000126123, + 0.005282300058752298, + 0.0038458683993667364, + 0.00255628046579659, + 0.004964854102581739, + 0.00315934419631958, + 0.0029648495838046074, + 0.0058496007695794106, + 0.003635213477537036, + 0.024850668385624886, + 0.022092236205935478, + 0.03225720673799515, + 0.020982839167118073, + 0.028138501569628716, + 0.028510628268122673, + 0.03777329623699188, + 0.03926457092165947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.006743056699633598 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188826739788055 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.018188826739788055, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188826739788055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09073932468891144, + 0.05150724574923515, + 0.09038510173559189, + 0.2446470558643341, + 0.20624400675296783, + 0.07723484933376312, + 0.07755827158689499, + 0.047876667231321335, + 0.03073136694729328, + 0.08307613432407379 + ], + "credit_map": [ + 0.0073141735047101974, + 0.006409074179828167, + 0.01556545402854681, + 0.009732433594763279, + 0.009288561530411243, + 0.022422196343541145, + 0.010496946983039379, + 0.009510485455393791, + 0.005774690303951502, + 0.0036091560032218695, + 0.0068266973830759525, + 0.0064078643918037415, + 0.006059257313609123, + 0.008498950861394405, + 0.006563629023730755, + 0.00776700209826231, + 0.008619760163128376, + 0.008884061127901077, + 0.012901157140731812, + 0.010388907045125961, + 0.011856256052851677, + 0.016021499410271645, + 0.006699484772980213, + 0.015013975091278553, + 0.03859204798936844, + 0.024055955931544304, + 0.03363936394453049, + 0.04185585677623749, + 0.0247510839253664, + 0.027400244027376175, + 0.02403431013226509, + 0.030318191275000572, + 0.022084034979343414, + 0.01600278913974762, + 0.02988792397081852, + 0.05119543895125389, + 0.011578594334423542, + 0.011001591570675373, + 0.012541157193481922, + 0.05195247381925583, + 0.011670454405248165, + 0.014077840372920036, + 0.01175826508551836, + 0.013972572982311249, + 0.005287510342895985, + 0.005593220237642527, + 0.007307651452720165, + 0.007567341439425945, + 0.008362256921827793, + 0.01229136809706688, + 0.009704568423330784, + 0.012361321598291397, + 0.013035637326538563, + 0.00824112817645073, + 0.0064232866279780865, + 0.007138701621443033, + 0.0074113826267421246, + 0.006491640582680702, + 0.006570732221007347, + 0.005182859022170305, + 0.003756729420274496, + 0.005991550628095865, + 0.006551812868565321, + 0.005919959861785173, + 0.003615004476159811, + 0.0035479681100696325, + 0.0036404954735189676, + 0.003979017958045006, + 0.00375005672685802, + 0.00400134501978755, + 0.004185396246612072, + 0.004012082703411579, + 0.009417020715773106, + 0.008276623673737049, + 0.012060578912496567, + 0.010426483117043972, + 0.008906270377337933, + 0.00999041274189949, + 0.010038148611783981, + 0.01396059338003397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.018188826739788055 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16031357645988464 + ], + "opposite_logprob_tokens": [ + -1.910313606262207 + ], + "logprob_reward": -0.16031357645988464, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16031357645988464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1110648363828659, + 0.15777413547039032, + 0.09371732920408249, + 0.08995779603719711, + 0.14610731601715088, + 0.04556174576282501, + 0.18881478905677795, + 0.07065317034721375, + 0.04235055670142174, + 0.05399833992123604 + ], + "credit_map": [ + 0.01265319436788559, + 0.009707536548376083, + 0.01764255203306675, + 0.009428275749087334, + 0.008657725527882576, + 0.020092826336622238, + 0.01995420642197132, + 0.01292851660400629, + 0.023093409836292267, + 0.016072968021035194, + 0.02184988558292389, + 0.016288062557578087, + 0.013275349512696266, + 0.021287189796566963, + 0.018739135935902596, + 0.027168137952685356, + 0.012495085597038269, + 0.00922796968370676, + 0.013211472891271114, + 0.009766001254320145, + 0.007357504218816757, + 0.010180533863604069, + 0.010783005505800247, + 0.020695755258202553, + 0.006127608008682728, + 0.006429486908018589, + 0.022148434072732925, + 0.012122691608965397, + 0.006187738850712776, + 0.012190325185656548, + 0.0071498313918709755, + 0.017601678147912025, + 0.016653044149279594, + 0.010455381125211716, + 0.021127920597791672, + 0.03383632376790047, + 0.010137725621461868, + 0.011251697316765785, + 0.009785722941160202, + 0.032859500497579575, + 0.004643276333808899, + 0.007189817726612091, + 0.0075980667024850845, + 0.0076287537813186646, + 0.0033822963014245033, + 0.0032778973691165447, + 0.005965657532215118, + 0.005875977221876383, + 0.0208571869879961, + 0.02142905257642269, + 0.01907973177731037, + 0.027064623311161995, + 0.021241778507828712, + 0.021680286154150963, + 0.020884282886981964, + 0.03657783567905426, + 0.012329771183431149, + 0.009754198603332043, + 0.008680847473442554, + 0.008740868419408798, + 0.005382409784942865, + 0.007117103319615126, + 0.011515282094478607, + 0.0071326931938529015, + 0.009019599296152592, + 0.006150023080408573, + 0.0041306461207568645, + 0.004938820842653513, + 0.003993519116193056, + 0.003941250499337912, + 0.005004765000194311, + 0.0051719327457249165, + 0.006059341132640839, + 0.006427412386983633, + 0.007396838627755642, + 0.006436162628233433, + 0.006942315027117729, + 0.006871160585433245, + 0.00800821278244257, + 0.005856897681951523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.16031357645988464 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, eye-level tracking shot follows a majestic hybrid creature as it gracefully soars through a dramatic sunset sky. The creature possesses the powerful, muscular body of a Tyrannosaurus Rex, covered in fine, interlocking scales of deep emerald green with a bronze sheen. Its mighty legs hug its body, and a long, heavy tail flows behind it, delicately adjusting its balance. The pterosaur's massive, leathery wings blend seamlessly into its powerful shoulders, their membranes a rich, earthy brown, revealing a delicate network of veins as the light filters through them. The creature glides with effortless grace, wings spread wide and steady, sailing through fluffy clouds soaked with the day's last rays. Their undersides glow a fiery orange, their crowns brushed with soft pinks and lavenders. A low-angled golden sun acts as a brilliant key light, capturing the sharp edges of each individual scale and creating a glowing halo around the creature's entire silhouette, imbuing the scene with a breathtaking, pristine tranquility.", + "seed": 18582, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_3/index_2264_seed_18582_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16162069085985423, + "vlm_logprob": -0.16162069085985423, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.16162069085985423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05806336551904678 + ], + "opposite_logprob_tokens": [ + -2.933063268661499 + ], + "logprob_reward": -0.05806336551904678, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05806336551904678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14225859940052032, + 0.18689179420471191, + 0.2091120183467865, + 0.08269989490509033, + 0.08791042119264603, + 0.029308728873729706, + 0.05270218476653099, + 0.03198133036494255, + 0.13471591472625732, + 0.042419157922267914 + ], + "credit_map": [ + 0.020473696291446686, + 0.009961326606571674, + 0.030972879379987717, + 0.011959944851696491, + 0.008846714161336422, + 0.012100180611014366, + 0.01454243529587984, + 0.03340141102671623, + 0.02510008029639721, + 0.0165819451212883, + 0.023975597694516182, + 0.04061564803123474, + 0.012454012408852577, + 0.0170090664178133, + 0.02835070714354515, + 0.022804705426096916, + 0.01515452191233635, + 0.030456043779850006, + 0.03599648177623749, + 0.017152555286884308, + 0.01577705889940262, + 0.01706521399319172, + 0.038479406386613846, + 0.03903072699904442, + 0.007336792536079884, + 0.009962846525013447, + 0.01023926306515932, + 0.01167826447635889, + 0.014316126704216003, + 0.007249005138874054, + 0.010746238753199577, + 0.011171352118253708, + 0.007923271507024765, + 0.010140222497284412, + 0.01137077808380127, + 0.011359070427715778, + 0.016486387699842453, + 0.006262425798922777, + 0.011916449293494225, + 0.012451810762286186, + 0.0033819861710071564, + 0.002758489456027746, + 0.003736815880984068, + 0.0042943996377289295, + 0.004885964095592499, + 0.002228604396805167, + 0.0029969150200486183, + 0.005025551188737154, + 0.012221431359648705, + 0.005400282330811024, + 0.007291923742741346, + 0.003659623209387064, + 0.006521393079310656, + 0.003488394897431135, + 0.0038269180804491043, + 0.010292206890881062, + 0.002983445767313242, + 0.002372210845351219, + 0.005639823619276285, + 0.006965161766856909, + 0.0035231688525527716, + 0.001913335407152772, + 0.003511870978400111, + 0.005072309169918299, + 0.008075973019003868, + 0.008808838203549385, + 0.02108681946992874, + 0.02637779526412487, + 0.010066267102956772, + 0.01376304030418396, + 0.014592833817005157, + 0.03194434195756912, + 0.0036879880353808403, + 0.003393450751900673, + 0.00807682704180479, + 0.00788307934999466, + 0.004829011857509613, + 0.006006087176501751, + 0.004269005265086889, + 0.004273705184459686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.05806336551904678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906803280115128 + ], + "opposite_logprob_tokens": [ + -2.4640679359436035 + ], + "logprob_reward": -0.08906803280115128, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906803280115128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46232113242149353, + 0.05672077462077141, + 0.1700289249420166, + 0.17099371552467346, + 0.01336507499217987, + 0.01833293028175831, + 0.020723465830087662, + 0.02487112022936344, + 0.047814857214689255, + 0.014828016981482506 + ], + "credit_map": [ + 0.01747320033609867, + 0.009263931773602962, + 0.17093291878700256, + 0.017653631046414375, + 0.007244231179356575, + 0.017219115048646927, + 0.028225138783454895, + 0.19430895149707794, + 0.005630750674754381, + 0.009343340992927551, + 0.007452461868524551, + 0.009344549849629402, + 0.003670768579468131, + 0.004942898638546467, + 0.008873812854290009, + 0.007462188135832548, + 0.013414507731795311, + 0.029714835807681084, + 0.026097005233168602, + 0.01661551743745804, + 0.01517198421061039, + 0.015129763633012772, + 0.031429268419742584, + 0.022456051781773567, + 0.010281004942953587, + 0.02040191926062107, + 0.019966768100857735, + 0.020657453685998917, + 0.03678949177265167, + 0.017796559259295464, + 0.02205790765583515, + 0.02304261364042759, + 0.0012739953817799687, + 0.0013500002678483725, + 0.0015739698428660631, + 0.0021189330145716667, + 0.0016179740196093917, + 0.0013018078170716763, + 0.0019415485439822078, + 0.002186845289543271, + 0.002112407935783267, + 0.0020279157906770706, + 0.0024150346871465445, + 0.00235719233751297, + 0.0033471696078777313, + 0.0016102202935144305, + 0.001654879655689001, + 0.0028081093914806843, + 0.00365129834972322, + 0.0016507146647199988, + 0.0028100309427827597, + 0.0014885533601045609, + 0.004447565414011478, + 0.0013081906363368034, + 0.0021306523121893406, + 0.0032364591024816036, + 0.001461970736272633, + 0.001656081760302186, + 0.004911830183118582, + 0.006937169004231691, + 0.0018750217277556658, + 0.0010413770796731114, + 0.0029832888394594193, + 0.004004380200058222, + 0.003491516225039959, + 0.0038651437498629093, + 0.0072227707132697105, + 0.009752608835697174, + 0.00524556590244174, + 0.004802050534635782, + 0.005442103836685419, + 0.007993098348379135, + 0.0011355134192854166, + 0.0013428892707452178, + 0.0021340528037399054, + 0.0022412543185055256, + 0.0022485856898128986, + 0.0010609915480017662, + 0.002517289947718382, + 0.0021474396344274282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.08906803280115128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763320829719305 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.00763320829719305, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763320829719305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19530698657035828, + 0.06651988625526428, + 0.14289414882659912, + 0.3724029064178467, + 0.02395622432231903, + 0.009927362203598022, + 0.011680157855153084, + 0.017945121973752975, + 0.15162332355976105, + 0.00774391321465373 + ], + "credit_map": [ + 0.008619396016001701, + 0.00445760041475296, + 0.070526123046875, + 0.00832789670675993, + 0.004289690405130386, + 0.0075970981270074844, + 0.011761635541915894, + 0.07972755283117294, + 0.007644436322152615, + 0.010743414983153343, + 0.008202144876122475, + 0.011003215797245502, + 0.004280958324670792, + 0.0064702825620770454, + 0.009823979809880257, + 0.008351457305252552, + 0.006606429815292358, + 0.008898708038032055, + 0.041240546852350235, + 0.010263043455779552, + 0.005080061499029398, + 0.013609329238533974, + 0.012435817159712315, + 0.044760219752788544, + 0.024123243987560272, + 0.04412631317973137, + 0.044772520661354065, + 0.0456710048019886, + 0.07790708541870117, + 0.03786236420273781, + 0.047995276749134064, + 0.04994508996605873, + 0.0018881247378885746, + 0.0025469926185905933, + 0.0030684228986501694, + 0.0035792167764157057, + 0.004178659990429878, + 0.0023340797051787376, + 0.003029411658644676, + 0.003331316402181983, + 0.0009362733107991517, + 0.0012297965586185455, + 0.001564444275572896, + 0.0012645627139136195, + 0.0011622186284512281, + 0.0006831241771578789, + 0.0010013120481744409, + 0.0020856307819485664, + 0.0008398506906814873, + 0.0007117640343494713, + 0.0017389251152053475, + 0.0010190814500674605, + 0.003483495442196727, + 0.0007780352025292814, + 0.0012484478065744042, + 0.0018605577060952783, + 0.0012366637820377946, + 0.0014711248222738504, + 0.0032398272305727005, + 0.004339503590017557, + 0.0018455007812008262, + 0.000901786086615175, + 0.0020258447621017694, + 0.0028848708607256413, + 0.00990508683025837, + 0.01102563925087452, + 0.023404525592923164, + 0.0303694698959589, + 0.013004109263420105, + 0.016273997724056244, + 0.01636168174445629, + 0.0312788151204586, + 0.0005255888099782169, + 0.0005465748836286366, + 0.0012159605976194143, + 0.0010728578781709075, + 0.0013604271225631237, + 0.000879761588294059, + 0.0011084213620051742, + 0.0010343212634325027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.00763320829719305 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581839606165886 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581839606165886, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581839606165886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6369831562042236, + 0.0444055050611496, + 0.03142080456018448, + 0.04041851684451103, + 0.031663499772548676, + 0.01652495563030243, + 0.024326378479599953, + 0.07725739479064941, + 0.07912948727607727, + 0.017870349809527397 + ], + "credit_map": [ + 0.02376805618405342, + 0.010848606005311012, + 0.23630687594413757, + 0.02502131648361683, + 0.010644149966537952, + 0.023602154105901718, + 0.038515087217092514, + 0.2682769000530243, + 0.004177704453468323, + 0.0034796178806573153, + 0.006799164693802595, + 0.005239734426140785, + 0.004323490429669619, + 0.004041118081659079, + 0.00977043155580759, + 0.006574244238436222, + 0.0022869945969432592, + 0.0026559950783848763, + 0.005770380608737469, + 0.0033521088771522045, + 0.003854314796626568, + 0.0030338962096720934, + 0.0037557436153292656, + 0.0067113726399838924, + 0.0028614643961191177, + 0.005229663569480181, + 0.006273411214351654, + 0.004691630601882935, + 0.006448063999414444, + 0.00508974539116025, + 0.004796765279024839, + 0.005027771927416325, + 0.0022315604146569967, + 0.0033432720229029655, + 0.003970989026129246, + 0.003015792928636074, + 0.006947749760001898, + 0.0035584699362516403, + 0.003937767818570137, + 0.004657898563891649, + 0.001488212961703539, + 0.0009174169972538948, + 0.0027506379410624504, + 0.0019096884643658996, + 0.0016479288460686803, + 0.000997512019239366, + 0.0026365327648818493, + 0.0041770245879888535, + 0.005382669623941183, + 0.0027323560789227486, + 0.0031398360151797533, + 0.0016300705028697848, + 0.0036567600909620523, + 0.0016189776360988617, + 0.002610958181321621, + 0.0035547506995499134, + 0.005979849025607109, + 0.0060426765121519566, + 0.013915074989199638, + 0.020616408437490463, + 0.0069136423990130424, + 0.003033972578123212, + 0.008495897054672241, + 0.012259871698915958, + 0.005749095231294632, + 0.005758855491876602, + 0.012336203828454018, + 0.01710236631333828, + 0.005280053708702326, + 0.008104274980723858, + 0.008218609727919102, + 0.01658003032207489, + 0.0016470893751829863, + 0.002516434993594885, + 0.002433760091662407, + 0.002643086016178131, + 0.002525229472666979, + 0.0018481116276234388, + 0.002002014545723796, + 0.002254624618217349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.020581839606165886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327570080757141 + ], + "opposite_logprob_tokens": [ + -0.7577570080757141 + ], + "logprob_reward": -0.6327570080757141, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327570080757141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5788803696632385, + 0.02153707854449749, + 0.19627471268177032, + 0.05186823010444641, + 0.013100509531795979, + 0.009889229200780392, + 0.009956657886505127, + 0.012041709385812283, + 0.08134735375642776, + 0.025104178115725517 + ], + "credit_map": [ + 0.022463219240307808, + 0.011807957664132118, + 0.21589361131191254, + 0.021559223532676697, + 0.008844384923577309, + 0.022009573876857758, + 0.03467412665486336, + 0.24162828922271729, + 0.0027735314797610044, + 0.0019107384141534567, + 0.0029277701396495104, + 0.002291701268404722, + 0.0021471937652677298, + 0.0016170431626960635, + 0.004055968020111322, + 0.0038131317123770714, + 0.010044884867966175, + 0.014358450658619404, + 0.04935481399297714, + 0.012953297235071659, + 0.0102365892380476, + 0.015425099059939384, + 0.02594919316470623, + 0.05795237049460411, + 0.003663973417133093, + 0.0062362318858504295, + 0.006214598193764687, + 0.006397770717740059, + 0.010339922271668911, + 0.005318748764693737, + 0.006706171669065952, + 0.0069908155128359795, + 0.0009350331383757293, + 0.001569868647493422, + 0.0017078053206205368, + 0.0018441297579556704, + 0.0021298001520335674, + 0.0008866747957654297, + 0.002037160564213991, + 0.001990037504583597, + 0.0011401866795495152, + 0.0007489316049031913, + 0.001580802840180695, + 0.0013939067721366882, + 0.0015970261301845312, + 0.0007032394642010331, + 0.0011811351869255304, + 0.0015440004644915462, + 0.0011654135305434465, + 0.0007827802328392863, + 0.0014698702143505216, + 0.000868802482727915, + 0.0017722033662721515, + 0.0008071010815910995, + 0.0011502858251333237, + 0.0019402008038014174, + 0.0012344710994511843, + 0.0009904516628012061, + 0.002136748982593417, + 0.002797403372824192, + 0.0008328420226462185, + 0.0007754489779472351, + 0.0014454302145168185, + 0.0018289136933162808, + 0.005785712506622076, + 0.006305173505097628, + 0.012214209884405136, + 0.014918696135282516, + 0.008526844903826714, + 0.008896428160369396, + 0.008618042804300785, + 0.016082242131233215, + 0.0020407666452229023, + 0.0033759111538529396, + 0.0032400020863860846, + 0.003206778783351183, + 0.003230971982702613, + 0.003721738699823618, + 0.0029353031422942877, + 0.003352705156430602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.6327570080757141 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, eye-level tracking shot follows a majestic hybrid creature as it gracefully soars through a dramatic sunset sky. The creature possesses the powerful, muscular body of a Tyrannosaurus Rex, covered in fine, interlocking scales of deep emerald green with a bronze sheen. Its mighty legs hug its body, and a long, heavy tail flows behind it, delicately adjusting its balance. The pterosaur's massive, leathery wings blend seamlessly into its powerful shoulders, their membranes a rich, earthy brown, revealing a delicate network of veins as the light filters through them. The creature glides with effortless grace, wings spread wide and steady, sailing through fluffy clouds soaked with the day's last rays. Their undersides glow a fiery orange, their crowns brushed with soft pinks and lavenders. A low-angled golden sun acts as a brilliant key light, capturing the sharp edges of each individual scale and creating a glowing halo around the creature's entire silhouette, imbuing the scene with a breathtaking, pristine tranquility.", + "seed": 18583, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_3/index_2264_seed_18583_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07227698424831033, + "vlm_logprob": -0.07227698424831033, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.07227698424831033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10532944649457932 + ], + "opposite_logprob_tokens": [ + -2.3553295135498047 + ], + "logprob_reward": -0.10532944649457932, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10532944649457932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08416467159986496, + 0.04221165552735329, + 0.05084560811519623, + 0.07177194207906723, + 0.05939220264554024, + 0.047709349542856216, + 0.1314578354358673, + 0.24915769696235657, + 0.031146708875894547, + 0.23214229941368103 + ], + "credit_map": [ + 0.006994401104748249, + 0.015453862957656384, + 0.006602080538868904, + 0.008107990026473999, + 0.01081550121307373, + 0.011873516254127026, + 0.010892785154283047, + 0.013424540869891644, + 0.0036576297134160995, + 0.0076403627172112465, + 0.0040528429672122, + 0.004735604394227266, + 0.004940480459481478, + 0.004738305229693651, + 0.005843822844326496, + 0.006602607201784849, + 0.004175703041255474, + 0.00657745311036706, + 0.005470469128340483, + 0.0046376134268939495, + 0.004628998693078756, + 0.007184648420661688, + 0.0062887780368328094, + 0.01188194565474987, + 0.011445675976574421, + 0.007507411763072014, + 0.011800509877502918, + 0.007453846745193005, + 0.00913071446120739, + 0.007427738513797522, + 0.008751722052693367, + 0.008254322223365307, + 0.0059736184775829315, + 0.004169559571892023, + 0.009206628426909447, + 0.006688205990940332, + 0.007803101558238268, + 0.009548462927341461, + 0.007650816813111305, + 0.008351806551218033, + 0.00783955492079258, + 0.004162895958870649, + 0.0046943859197199345, + 0.0039938073605299, + 0.005079376045614481, + 0.007084676064550877, + 0.00753084197640419, + 0.00732381222769618, + 0.01624009758234024, + 0.016486575827002525, + 0.012417307123541832, + 0.013187102973461151, + 0.03317456319928169, + 0.015099575743079185, + 0.012817446142435074, + 0.012035161256790161, + 0.0976155698299408, + 0.021749839186668396, + 0.09493494778871536, + 0.007448364980518818, + 0.006864818744361401, + 0.004266216419637203, + 0.011930054984986782, + 0.004347912967205048, + 0.004971408750861883, + 0.005474523641169071, + 0.004298354499042034, + 0.0039568557403981686, + 0.002688703825697303, + 0.0036993371322751045, + 0.002381131751462817, + 0.0036763909738510847, + 0.045118529349565506, + 0.04509152099490166, + 0.031501952558755875, + 0.02998974733054638, + 0.025264166295528412, + 0.015359361656010151, + 0.023059066385030746, + 0.016757972538471222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.10532944649457932 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043110329657793045 + ], + "opposite_logprob_tokens": [ + -3.1681103706359863 + ], + "logprob_reward": -0.043110329657793045, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043110329657793045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07742077112197876, + 0.05318043753504753, + 0.030524631962180138, + 0.14038316905498505, + 0.04625805467367172, + 0.044678036123514175, + 0.23982711136341095, + 0.2610235810279846, + 0.039245523512363434, + 0.06745868921279907 + ], + "credit_map": [ + 0.005760353058576584, + 0.013799426145851612, + 0.00772731751203537, + 0.007137025240808725, + 0.009032953530550003, + 0.010630885139107704, + 0.009793978184461594, + 0.013538840226829052, + 0.004083996172994375, + 0.006529574748128653, + 0.003595274407416582, + 0.005849483888596296, + 0.004982895217835903, + 0.010490207932889462, + 0.0057466719299554825, + 0.011902334168553352, + 0.003203130094334483, + 0.0044728913344442844, + 0.0026572286151349545, + 0.0034764856100082397, + 0.003409868571907282, + 0.003916454501450062, + 0.003543724538758397, + 0.005844848230481148, + 0.021551407873630524, + 0.015079081058502197, + 0.01815248280763626, + 0.017238643020391464, + 0.022915193811058998, + 0.013779670000076294, + 0.01618165709078312, + 0.015485024079680443, + 0.005163692869246006, + 0.003706492716446519, + 0.009310086257755756, + 0.004928229842334986, + 0.005651210900396109, + 0.003591034794226289, + 0.007077168673276901, + 0.006830137688666582, + 0.00787147507071495, + 0.00381701672449708, + 0.004487289115786552, + 0.003610871033743024, + 0.0048990994691848755, + 0.00658646784722805, + 0.006844316143542528, + 0.0065615009516477585, + 0.03188567981123924, + 0.032891035079956055, + 0.027372604236006737, + 0.026197608560323715, + 0.05159005522727966, + 0.023130755871534348, + 0.023605963215231895, + 0.023153390735387802, + 0.10774485766887665, + 0.00952920876443386, + 0.10786329209804535, + 0.006137173622846603, + 0.0060860589146614075, + 0.0040551163256168365, + 0.015217798762023449, + 0.004390063229948282, + 0.008044408634305, + 0.00666919955983758, + 0.0069998181425035, + 0.00418900977820158, + 0.003447216935455799, + 0.0038373672869056463, + 0.002489556558430195, + 0.0035689459182322025, + 0.011030073277652264, + 0.009139541536569595, + 0.006924275308847427, + 0.00785845797508955, + 0.00903000496327877, + 0.006733126938343048, + 0.00915546715259552, + 0.007587745785713196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.043110329657793045 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744359154254198 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.006744359154254198, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744359154254198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12059783935546875, + 0.05382160469889641, + 0.05984380841255188, + 0.13090188801288605, + 0.052961841225624084, + 0.058448001742362976, + 0.084275983273983, + 0.2542670667171478, + 0.05366554111242294, + 0.13121642172336578 + ], + "credit_map": [ + 0.007507037371397018, + 0.02232670597732067, + 0.010682955384254456, + 0.010139905847609043, + 0.01415034756064415, + 0.017310000956058502, + 0.015669193118810654, + 0.022811701521277428, + 0.006565281189978123, + 0.011398999951779842, + 0.004440949764102697, + 0.004581228829920292, + 0.00521125691011548, + 0.006687619257718325, + 0.006619746796786785, + 0.008316523395478725, + 0.007287015672773123, + 0.005659725982695818, + 0.005374655593186617, + 0.0042328378185629845, + 0.00460624136030674, + 0.011697390116751194, + 0.007130897603929043, + 0.01385504100471735, + 0.016095424070954323, + 0.013801989145576954, + 0.02100060135126114, + 0.014817388728260994, + 0.014224419370293617, + 0.012991671450436115, + 0.015254098922014236, + 0.02271628938615322, + 0.006287539843469858, + 0.00505592068657279, + 0.010613271035254002, + 0.005776611156761646, + 0.006997919175773859, + 0.003055885899811983, + 0.007738552987575531, + 0.007436139974743128, + 0.009804587811231613, + 0.00516595970839262, + 0.006280047819018364, + 0.005163037218153477, + 0.005789171904325485, + 0.009297593496739864, + 0.008269506506621838, + 0.008678095415234566, + 0.010542873293161392, + 0.00881174299865961, + 0.009386628866195679, + 0.009239940904080868, + 0.019359655678272247, + 0.009236014448106289, + 0.009816545993089676, + 0.007882577367126942, + 0.10150548815727234, + 0.012029689736664295, + 0.10162895917892456, + 0.008011480793356895, + 0.006971132010221481, + 0.004334812052547932, + 0.014668934978544712, + 0.005116571206599474, + 0.010403086431324482, + 0.010161316022276878, + 0.009335355833172798, + 0.005895745009183884, + 0.003961401060223579, + 0.005133211147040129, + 0.0040454138070344925, + 0.004730012267827988, + 0.0264178067445755, + 0.02749774046242237, + 0.021929988637566566, + 0.01725504733622074, + 0.012083965353667736, + 0.007349231746047735, + 0.00992571096867323, + 0.00875694677233696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.006744359154254198 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348866522312164 + ], + "opposite_logprob_tokens": [ + -3.651348829269409 + ], + "logprob_reward": -0.026348866522312164, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348866522312164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1803608536720276, + 0.03884003311395645, + 0.06731592118740082, + 0.18565477430820465, + 0.0556405670940876, + 0.07092618197202682, + 0.08723099529743195, + 0.0665360540151596, + 0.08806058764457703, + 0.1594340205192566 + ], + "credit_map": [ + 0.013776592910289764, + 0.042903341352939606, + 0.012876152992248535, + 0.017499981448054314, + 0.02049182914197445, + 0.023176971822977066, + 0.016376536339521408, + 0.033259447664022446, + 0.003941683564335108, + 0.0031967677641659975, + 0.004993905313313007, + 0.004541753325611353, + 0.0039061973802745342, + 0.005369060672819614, + 0.004964946303516626, + 0.007925722748041153, + 0.006303588859736919, + 0.007799587678164244, + 0.014558817259967327, + 0.007008703891187906, + 0.004618208855390549, + 0.009056399576365948, + 0.00788246188312769, + 0.010088159702718258, + 0.012839124538004398, + 0.008141952566802502, + 0.02311524748802185, + 0.011339973658323288, + 0.011214102618396282, + 0.026083607226610184, + 0.025602417066693306, + 0.06731835752725601, + 0.005982611794024706, + 0.004676606506109238, + 0.008844112977385521, + 0.004982423968613148, + 0.004982935264706612, + 0.005405437666922808, + 0.009960579685866833, + 0.010805861093103886, + 0.009423797018826008, + 0.0064679645001888275, + 0.013677572831511497, + 0.007021764758974314, + 0.007444728631526232, + 0.007929584942758083, + 0.01022451464086771, + 0.008736255578696728, + 0.012322313152253628, + 0.012016634456813335, + 0.009827034547924995, + 0.009905707091093063, + 0.0162197332829237, + 0.008975962176918983, + 0.008832408115267754, + 0.009131206199526787, + 0.013338909484446049, + 0.010994741693139076, + 0.012232706882059574, + 0.009095131419599056, + 0.0063649555668234825, + 0.00446686428040266, + 0.005710049998015165, + 0.004332696087658405, + 0.021127481013536453, + 0.02060236968100071, + 0.011469507589936256, + 0.010194476693868637, + 0.005777022801339626, + 0.006962113548070192, + 0.004452270921319723, + 0.007475349120795727, + 0.03341793268918991, + 0.03470943123102188, + 0.03325827047228813, + 0.019264642149209976, + 0.01177644357085228, + 0.00760310934856534, + 0.00875723548233509, + 0.010646956972777843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.026348866522312164 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985191941261292 + ], + "opposite_logprob_tokens": [ + -1.8048518896102905 + ], + "logprob_reward": -0.17985191941261292, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985191941261292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13996608555316925, + 0.17624424397945404, + 0.05749129503965378, + 0.04105796664953232, + 0.0690024346113205, + 0.10393594950437546, + 0.07994043827056885, + 0.2549290657043457, + 0.023689022287726402, + 0.05374348908662796 + ], + "credit_map": [ + 0.0166278388351202, + 0.024390455335378647, + 0.009898319840431213, + 0.01765182614326477, + 0.015296027995646, + 0.017919091507792473, + 0.018504714593291283, + 0.019677812233567238, + 0.017007730901241302, + 0.01566387712955475, + 0.013264666311442852, + 0.020403582602739334, + 0.015316163189709187, + 0.036806363612413406, + 0.018834620714187622, + 0.03894723951816559, + 0.006525903474539518, + 0.007476126309484243, + 0.004616414196789265, + 0.007093184627592564, + 0.007374405395239592, + 0.0069502671249210835, + 0.008463188074529171, + 0.008991803042590618, + 0.006287675816565752, + 0.004124358296394348, + 0.006494835484772921, + 0.003240549238398671, + 0.004523014649748802, + 0.003750225529074669, + 0.0049440450966358185, + 0.007693259511142969, + 0.007263317704200745, + 0.00499739358201623, + 0.008729089982807636, + 0.0074391113594174385, + 0.0086458008736372, + 0.01273480523377657, + 0.006839591544121504, + 0.012353330850601196, + 0.015307821333408356, + 0.008146317675709724, + 0.01162714697420597, + 0.009130553342401981, + 0.013632198795676231, + 0.014249039813876152, + 0.017114005982875824, + 0.014728865586221218, + 0.009678689762949944, + 0.011129535734653473, + 0.009011304937303066, + 0.009461726993322372, + 0.014615987427532673, + 0.009711029008030891, + 0.008886349387466908, + 0.00744581688195467, + 0.08102878928184509, + 0.03314701467752457, + 0.08397796750068665, + 0.012917078100144863, + 0.011692067608237267, + 0.008987528271973133, + 0.016713999211788177, + 0.006464633159339428, + 0.004404950421303511, + 0.005242123268544674, + 0.003186420537531376, + 0.002896923338994384, + 0.0017187088960781693, + 0.0022690535988658667, + 0.0012185018276795745, + 0.0027523417957127094, + 0.007976115681231022, + 0.006000993773341179, + 0.007832609117031097, + 0.007266531698405743, + 0.006102405488491058, + 0.006000063847750425, + 0.007719111163169146, + 0.0048456634394824505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.17985191941261292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid, eye-level tracking shot follows a majestic hybrid creature as it gracefully soars through a dramatic sunset sky. The creature possesses the powerful, muscular body of a Tyrannosaurus Rex, covered in fine, interlocking scales of deep emerald green with a bronze sheen. Its mighty legs hug its body, and a long, heavy tail flows behind it, delicately adjusting its balance. The pterosaur's massive, leathery wings blend seamlessly into its powerful shoulders, their membranes a rich, earthy brown, revealing a delicate network of veins as the light filters through them. The creature glides with effortless grace, wings spread wide and steady, sailing through fluffy clouds soaked with the day's last rays. Their undersides glow a fiery orange, their crowns brushed with soft pinks and lavenders. A low-angled golden sun acts as a brilliant key light, capturing the sharp edges of each individual scale and creating a glowing halo around the creature's entire silhouette, imbuing the scene with a breathtaking, pristine tranquility.", + "seed": 18584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_3/index_2264_seed_18584_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.055682362616062166, + "vlm_logprob": -0.055682362616062166, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.055682362616062166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025811947882175446 + ], + "opposite_logprob_tokens": [ + -3.9008119106292725 + ], + "logprob_reward": -0.025811947882175446, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025811947882175446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22594797611236572, + 0.06861451268196106, + 0.12438406050205231, + 0.10543838143348694, + 0.18564492464065552, + 0.04722342640161514, + 0.040640488266944885, + 0.0868089348077774, + 0.023474175482988358, + 0.09182308614253998 + ], + "credit_map": [ + 0.02118058316409588, + 0.020872894674539566, + 0.010080981999635696, + 0.007879910059273243, + 0.06539489328861237, + 0.08000025898218155, + 0.010601988062262535, + 0.009936473332345486, + 0.010765261016786098, + 0.007730345241725445, + 0.009082388132810593, + 0.008011539466679096, + 0.009487718343734741, + 0.010563981719315052, + 0.00641895504668355, + 0.006554328836500645, + 0.014725040644407272, + 0.013280417770147324, + 0.02040833607316017, + 0.010901050642132759, + 0.01996375061571598, + 0.00936469342559576, + 0.012076765298843384, + 0.023664012551307678, + 0.012135659344494343, + 0.01923449896275997, + 0.013949528336524963, + 0.01960204914212227, + 0.012078829109668732, + 0.009724311530590057, + 0.010112997144460678, + 0.008600513450801373, + 0.014508141204714775, + 0.020829739049077034, + 0.020711654797196388, + 0.02670459635555744, + 0.027247752994298935, + 0.026049859821796417, + 0.022997310385107994, + 0.02659587375819683, + 0.004794031381607056, + 0.005075192078948021, + 0.006154799833893776, + 0.0094620231539011, + 0.006700210738927126, + 0.006582924630492926, + 0.0038719428703188896, + 0.004582306370139122, + 0.005551477428525686, + 0.004893904086202383, + 0.006331963464617729, + 0.006734469439834356, + 0.004063477274030447, + 0.0033727670088410378, + 0.003890599124133587, + 0.005801835563033819, + 0.010520062409341335, + 0.01073961891233921, + 0.011183512397110462, + 0.007242383901029825, + 0.013583715073764324, + 0.01301480084657669, + 0.009804603643715382, + 0.010720249265432358, + 0.0027493364177644253, + 0.0028969442937523127, + 0.0037634491454809904, + 0.003652521874755621, + 0.003327444428578019, + 0.0027839806862175465, + 0.0018135776044800878, + 0.002486921614035964, + 0.007141970098018646, + 0.01240106113255024, + 0.007699496578425169, + 0.00929335132241249, + 0.01193180587142706, + 0.027950428426265717, + 0.009019734337925911, + 0.00638524629175663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.025811947882175446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026369880884885788 + ], + "opposite_logprob_tokens": [ + -3.65136981010437 + ], + "logprob_reward": -0.026369880884885788, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026369880884885788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2608931362628937, + 0.10332456976175308, + 0.13648775219917297, + 0.18201734125614166, + 0.06237264350056648, + 0.03606019914150238, + 0.03708085045218468, + 0.08085432648658752, + 0.02721172198653221, + 0.07369740307331085 + ], + "credit_map": [ + 0.034693680703639984, + 0.021821068599820137, + 0.026935994625091553, + 0.017266428098082542, + 0.052439551800489426, + 0.05697737634181976, + 0.019689347594976425, + 0.03106973133981228, + 0.01365804485976696, + 0.013526445254683495, + 0.017255352810025215, + 0.011939015239477158, + 0.011431412771344185, + 0.016766753047704697, + 0.008051403798162937, + 0.010696145705878735, + 0.015262654051184654, + 0.03516729176044464, + 0.016191812232136726, + 0.01034523919224739, + 0.015460600145161152, + 0.018199928104877472, + 0.014757255092263222, + 0.011102982796728611, + 0.020810957998037338, + 0.019192036241292953, + 0.028634624555706978, + 0.038670457899570465, + 0.021657083183526993, + 0.011016028001904488, + 0.01981058530509472, + 0.022225582972168922, + 0.004858602769672871, + 0.008227774873375893, + 0.006959342397749424, + 0.010748169384896755, + 0.015074613504111767, + 0.004102199804037809, + 0.0050399284809827805, + 0.007362015079706907, + 0.004996823146939278, + 0.005335942376405001, + 0.0056418548338115215, + 0.005751552525907755, + 0.0050864554941654205, + 0.003119400469586253, + 0.0030647891107946634, + 0.003063386771827936, + 0.004644735250622034, + 0.005212165415287018, + 0.004314083606004715, + 0.0047734626568853855, + 0.005713616497814655, + 0.0038301930762827396, + 0.004201987758278847, + 0.004390610847622156, + 0.008744945749640465, + 0.00983105506747961, + 0.009532829746603966, + 0.005870409309864044, + 0.013603419065475464, + 0.01551845297217369, + 0.008481694385409355, + 0.009271521121263504, + 0.002974569099023938, + 0.003359355963766575, + 0.005845696199685335, + 0.0036560550797730684, + 0.002814888022840023, + 0.0035177920944988728, + 0.0024450733326375484, + 0.0025982961524277925, + 0.007405684795230627, + 0.008743747137486935, + 0.010165812447667122, + 0.009931590408086777, + 0.008019431494176388, + 0.020087450742721558, + 0.005165118724107742, + 0.004178570583462715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.026369880884885788 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086426492780447 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.0086426492780447, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086426492780447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48753660917282104, + 0.09312758594751358, + 0.07264220714569092, + 0.04907873645424843, + 0.08952387422323227, + 0.03184979408979416, + 0.043668005615472794, + 0.05905784294009209, + 0.020833242684602737, + 0.052682120352983475 + ], + "credit_map": [ + 0.043559182435274124, + 0.04830830171704292, + 0.03870031610131264, + 0.018750198185443878, + 0.12776660919189453, + 0.15049989521503448, + 0.024177879095077515, + 0.03577423095703125, + 0.01138249970972538, + 0.008749942295253277, + 0.016198519617319107, + 0.010272108018398285, + 0.015381532721221447, + 0.012877805158495903, + 0.008571157231926918, + 0.009694021195173264, + 0.007474674377590418, + 0.018288150429725647, + 0.008475358597934246, + 0.006176317110657692, + 0.005898552946746349, + 0.009554564952850342, + 0.008601408451795578, + 0.008173178881406784, + 0.005720910150557756, + 0.0076630571857094765, + 0.006052357144653797, + 0.007958597503602505, + 0.006031080149114132, + 0.0037411570083349943, + 0.005350395105779171, + 0.006561181973665953, + 0.007654484361410141, + 0.01142258569598198, + 0.01155777182430029, + 0.016846604645252228, + 0.009527130983769894, + 0.010165782645344734, + 0.009673422202467918, + 0.012676089070737362, + 0.003649319987744093, + 0.0044118971563875675, + 0.004433698486536741, + 0.00429428368806839, + 0.006552729289978743, + 0.002511933445930481, + 0.003479844192042947, + 0.0025160866789519787, + 0.006618548184633255, + 0.005816914606839418, + 0.006014482583850622, + 0.0046322960406541824, + 0.004407858941704035, + 0.003509435337036848, + 0.004567301366478205, + 0.008101169019937515, + 0.0059503475204110146, + 0.007199421525001526, + 0.00737167289480567, + 0.005180459935218096, + 0.009923395700752735, + 0.009521420113742352, + 0.006375156342983246, + 0.007535974960774183, + 0.002649364760145545, + 0.0025734652299433947, + 0.003944285213947296, + 0.003718018066138029, + 0.0021416463423520327, + 0.0022489798720926046, + 0.0018008517799898982, + 0.0017566310707479715, + 0.005426906049251556, + 0.0061357878148555756, + 0.007364691700786352, + 0.007359989918768406, + 0.006868987809866667, + 0.011691759340465069, + 0.004134215880185366, + 0.0036997832357883453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0086426492780447 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607118360698223 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.01607118360698223, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607118360698223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14969633519649506, + 0.0862894058227539, + 0.21028336882591248, + 0.08464615792036057, + 0.07742206752300262, + 0.1308894008398056, + 0.07441452145576477, + 0.0630098357796669, + 0.028912866488099098, + 0.09443604946136475 + ], + "credit_map": [ + 0.01862618327140808, + 0.019432011991739273, + 0.011188591830432415, + 0.009410958737134933, + 0.032329633831977844, + 0.03580373898148537, + 0.010521771386265755, + 0.012383452616631985, + 0.011990637518465519, + 0.008053814992308617, + 0.012967019341886044, + 0.010611389763653278, + 0.014177989214658737, + 0.013656909577548504, + 0.006244008895009756, + 0.008587638847529888, + 0.024419942870736122, + 0.017894215881824493, + 0.037481699138879776, + 0.01772882603108883, + 0.03334229812026024, + 0.014575066044926643, + 0.021536946296691895, + 0.04330437257885933, + 0.01256032008677721, + 0.008995323441922665, + 0.014775355346500874, + 0.021132243797183037, + 0.005973500665277243, + 0.006593537051230669, + 0.0069953566417098045, + 0.007620526477694511, + 0.010097537189722061, + 0.008421543054282665, + 0.009848050773143768, + 0.012241040356457233, + 0.012269160710275173, + 0.00603656517341733, + 0.006279642693698406, + 0.012228532694280148, + 0.018976623192429543, + 0.01803618296980858, + 0.022840995341539383, + 0.024647220969200134, + 0.012409781105816364, + 0.009061255492269993, + 0.014166788198053837, + 0.010750556364655495, + 0.01056958269327879, + 0.01089102029800415, + 0.009919848293066025, + 0.010067716240882874, + 0.007920845411717892, + 0.005880290176719427, + 0.008744620718061924, + 0.01042060274630785, + 0.007926796562969685, + 0.007024927530437708, + 0.01260219793766737, + 0.006216063164174557, + 0.006790907122194767, + 0.0060253324918448925, + 0.008355078287422657, + 0.00806853361427784, + 0.004584132228046656, + 0.003685656702145934, + 0.004406161140650511, + 0.0046739461831748486, + 0.003481852589175105, + 0.0025617522187530994, + 0.0028485043440014124, + 0.002670860616490245, + 0.009359861724078655, + 0.013391018845140934, + 0.009525207802653313, + 0.012562131509184837, + 0.018794948235154152, + 0.013059078715741634, + 0.010255765169858932, + 0.007488030940294266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.01607118360698223 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hybrid creature combining the body of a Tyrannosaurus Rex with pterosaur wings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the creature's body covered in fine, interlocking scales of deep emerald green with a bronze sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creature's wings have rich, earthy brown membranes with a visible network of veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the creature gliding through a sunset sky with fluffy clouds, showing fiery orange undersides and pink/lavender crowns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot an eye-level tracking shot following the creature as it soars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20151615142822266 + ], + "opposite_logprob_tokens": [ + -1.7015161514282227 + ], + "logprob_reward": -0.20151615142822266, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20151615142822266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1937604397535324, + 0.10808563232421875, + 0.10103640705347061, + 0.13845019042491913, + 0.06660139560699463, + 0.062216617166996, + 0.09439501911401749, + 0.13020144402980804, + 0.05360023304820061, + 0.05165265500545502 + ], + "credit_map": [ + 0.033947933465242386, + 0.023650765419006348, + 0.01965387910604477, + 0.015942562371492386, + 0.03395858407020569, + 0.026612630113959312, + 0.02322501689195633, + 0.016769085079431534, + 0.014784354716539383, + 0.01367878820747137, + 0.015710998326539993, + 0.011066360399127007, + 0.00999334268271923, + 0.020912613719701767, + 0.009254694916307926, + 0.012684478424489498, + 0.013408870436251163, + 0.011636669747531414, + 0.016449855640530586, + 0.010051676072180271, + 0.014494801871478558, + 0.007343767210841179, + 0.009799860417842865, + 0.01785091497004032, + 0.01681622862815857, + 0.019217604771256447, + 0.01977059431374073, + 0.02882237359881401, + 0.014225908555090427, + 0.01181080099195242, + 0.01503591239452362, + 0.012750756926834583, + 0.006706953514367342, + 0.006655159872025251, + 0.007536249700933695, + 0.010117664001882076, + 0.009568148292601109, + 0.00973997451364994, + 0.007289317436516285, + 0.00898792315274477, + 0.007554531563073397, + 0.005767229478806257, + 0.007944478653371334, + 0.011025388725101948, + 0.011326618492603302, + 0.004723170306533575, + 0.007951507344841957, + 0.005923693999648094, + 0.014689581468701363, + 0.014526992104947567, + 0.012037944979965687, + 0.01161256618797779, + 0.00812133215367794, + 0.007963918149471283, + 0.009403783828020096, + 0.016038905829191208, + 0.013559571467339993, + 0.016115667298436165, + 0.01725412905216217, + 0.013717692345380783, + 0.01597757823765278, + 0.020130256190896034, + 0.01623130589723587, + 0.017215251922607422, + 0.008401535451412201, + 0.006961406674236059, + 0.010705960914492607, + 0.005825006403028965, + 0.005593059118837118, + 0.005734575912356377, + 0.005050529260188341, + 0.005328156054019928, + 0.004826128017157316, + 0.007055783644318581, + 0.005977288354188204, + 0.006128509528934956, + 0.007486504036933184, + 0.012626861222088337, + 0.004226588644087315, + 0.0033249908592551947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.20151615142822266 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_4/index_tensor([3628, 3628])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_4/index_tensor([3628, 3628])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3b320a08417860c22d5b20d8979114a29387930b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_4/index_tensor([3628, 3628])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An eye-level close-up, capturing the subtle nervous tremor of a handheld camera, reveals a massive great white shark suspended, motionless, in the vast abyss. The shark's massive, scarred body, a gradient of dark slate gray on top and a distinctly white underbelly, fills the frame, its skin textured with the subtle roughness of dense dermis. Surrounded by an endless expanse of deep cobalt blue, faint, ethereal sunlight penetrates the water from above, catching flecks of marine snow and nearly illuminating its dorsal flank before being swallowed by the overwhelming darkness below. Its body is a masterpiece of predatory design, held in eerie stillness, poised for an imminent, explosive attack. Its left eye, a black, unblinking, soulless orb, is fixed directly on the lens, creating an unsettling and direct confrontation, while its jaw, slightly parted, reveals rows of jagged, triangular teeth. The only perceptible movement is a slight, eerie tremor of the camera and the slow, rhythmic opening of the shark's gills, creating a chilling atmosphere of pure, breathless tension and primal fear.", + "seed": 3932, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_4/index_3628_seed_3932_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.041771398205310104, + "vlm_logprob": -0.041771398205310104, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.041771398205310104, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027942240238189697 + ], + "opposite_logprob_tokens": [ + -4.027942180633545 + ], + "logprob_reward": -0.027942240238189697, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.027942240238189697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5226741433143616, + 0.26916033029556274, + 0.10960586369037628, + 0.020445412024855614, + 0.015652304515242577, + 0.005701324436813593, + 0.01598801091313362, + 0.006603738758713007, + 0.009361357428133488, + 0.024807479232549667 + ], + "credit_map": [ + 0.0624675489962101, + 0.023512734100222588, + 0.02009906992316246, + 0.009965511970221996, + 0.07146131992340088, + 0.12677446007728577, + 0.02274917997419834, + 0.18564435839653015, + 0.03742643818259239, + 0.01563732698559761, + 0.010944800451397896, + 0.008012653328478336, + 0.030294833704829216, + 0.06102491542696953, + 0.014121351763606071, + 0.09169802069664001, + 0.01468336209654808, + 0.0071248337626457214, + 0.0046584727242589, + 0.004875733982771635, + 0.01240081898868084, + 0.02723568119108677, + 0.00613112049177289, + 0.03249584138393402, + 0.0019399180309846997, + 0.0026977260131388903, + 0.0014619085704907775, + 0.0016374454135075212, + 0.005200990010052919, + 0.0034226304851472378, + 0.0018616901943460107, + 0.002223104005679488, + 0.0016852596309036016, + 0.001670239376835525, + 0.001873877365142107, + 0.0017907342407852411, + 0.002942556282505393, + 0.0025156354531645775, + 0.0013679901603609324, + 0.0018060131696984172, + 0.0007651119376532733, + 0.0006237280904315412, + 0.0005406601121649146, + 0.0005299879121594131, + 0.001021991833113134, + 0.0009056847193278372, + 0.0006558797904290259, + 0.00065828044898808, + 0.0018611847190186381, + 0.002486202400177717, + 0.0014674216508865356, + 0.0012036709813401103, + 0.0036377348005771637, + 0.002388615859672427, + 0.0012891185469925404, + 0.0016540635842829943, + 0.0007953702588565648, + 0.0007641988340765238, + 0.0008985510794445872, + 0.0005256263539195061, + 0.0009089685045182705, + 0.0009364339057356119, + 0.0007538176141679287, + 0.001020772848278284, + 0.0012488476932048798, + 0.0017021672101691365, + 0.001165332505479455, + 0.000692827336024493, + 0.0011720909969881177, + 0.0011897666845470667, + 0.001121502136811614, + 0.0010688243201002479, + 0.002198094967752695, + 0.006560577545315027, + 0.002289347117766738, + 0.001588807557709515, + 0.0033803354017436504, + 0.0020470612216740847, + 0.0021112922113388777, + 0.004631963558495045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.027942240238189697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014212106354534626 + ], + "opposite_logprob_tokens": [ + -4.264212131500244 + ], + "logprob_reward": -0.014212106354534626, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014212106354534626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2862137258052826, + 0.3434818685054779, + 0.20474369823932648, + 0.027675596997141838, + 0.0296791885048151, + 0.022631805390119553, + 0.03337566927075386, + 0.012572546489536762, + 0.01707479916512966, + 0.022551104426383972 + ], + "credit_map": [ + 0.03529805690050125, + 0.013721052557229996, + 0.01151325460523367, + 0.005294069182127714, + 0.04046158120036125, + 0.06561500579118729, + 0.013871371746063232, + 0.10043934732675552, + 0.0469156950712204, + 0.019504249095916748, + 0.01421255525201559, + 0.011222302913665771, + 0.040075935423374176, + 0.07784069329500198, + 0.01917256787419319, + 0.11453787982463837, + 0.02652556262910366, + 0.014080001041293144, + 0.009336911141872406, + 0.00856911949813366, + 0.023043692111968994, + 0.05097052827477455, + 0.011646797880530357, + 0.06057107076048851, + 0.0027018149849027395, + 0.003184103174135089, + 0.0024962639436125755, + 0.003482665866613388, + 0.004893424920737743, + 0.004202671814709902, + 0.0024916641414165497, + 0.004222987685352564, + 0.004262466449290514, + 0.002702881582081318, + 0.003959428984671831, + 0.003790434915572405, + 0.0044632297940552235, + 0.004000798799097538, + 0.0029723444022238255, + 0.0035276056732982397, + 0.002917157020419836, + 0.0021916073746979237, + 0.0016556453192606568, + 0.0016464375657960773, + 0.0028029256500303745, + 0.0036693348083645105, + 0.0019914780277758837, + 0.005757220089435577, + 0.003909924533218145, + 0.0036852858029305935, + 0.002493771491572261, + 0.0021585721988230944, + 0.009196142666041851, + 0.0059160152450203896, + 0.0025964942760765553, + 0.0034194604959338903, + 0.0017025682609528303, + 0.0014287977246567607, + 0.0022265564184635878, + 0.0008571045473217964, + 0.0011809773277491331, + 0.001854973379522562, + 0.0010773183312267065, + 0.0022442510817199945, + 0.002059690188616514, + 0.002560883527621627, + 0.0018355704378336668, + 0.0013617119984701276, + 0.003115015337243676, + 0.00245237210765481, + 0.001880517229437828, + 0.0018090375233441591, + 0.0020486917346715927, + 0.007295833434909582, + 0.0017340820049867034, + 0.0013615225907415152, + 0.002261850517243147, + 0.0018632858991622925, + 0.0015349527820944786, + 0.00445088604465127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.014212106354534626 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948459263890982 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948459263890982, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948459263890982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25592339038848877, + 0.5941333174705505, + 0.031042858958244324, + 0.009255921468138695, + 0.0196906216442585, + 0.03217211738228798, + 0.01433586236089468, + 0.010556154884397984, + 0.016424110159277916, + 0.01646559126675129 + ], + "credit_map": [ + 0.029651392251253128, + 0.010068911127746105, + 0.009437613189220428, + 0.004564345348626375, + 0.027502015233039856, + 0.07160738855600357, + 0.010004290379583836, + 0.09308742731809616, + 0.08254969120025635, + 0.03565488010644913, + 0.02442443184554577, + 0.018630903214216232, + 0.06624112278223038, + 0.13467274606227875, + 0.03161033242940903, + 0.20034922659397125, + 0.0035092851612716913, + 0.0031084215734153986, + 0.001349723432213068, + 0.0015829192707315087, + 0.005932216998189688, + 0.006447494029998779, + 0.0019167560385540128, + 0.007196043152362108, + 0.0010921650100499392, + 0.001612833235412836, + 0.0006873536622151732, + 0.0008356840698979795, + 0.0013510819990187883, + 0.0014498310629278421, + 0.0008691927068866789, + 0.0013577797217294574, + 0.0024478984996676445, + 0.002017807913944125, + 0.0025929890107363462, + 0.0024545262567698956, + 0.002785708988085389, + 0.0027624741196632385, + 0.0019524643430486321, + 0.0026767526287585497, + 0.006789305713027716, + 0.002736336085945368, + 0.0018660433124750853, + 0.002018126891925931, + 0.005189751740545034, + 0.005403508432209492, + 0.0023786071687936783, + 0.005790439434349537, + 0.002084681997075677, + 0.0012068470241501927, + 0.0011412135791033506, + 0.0009112635743804276, + 0.0023474611807614565, + 0.002682264195755124, + 0.001266764709725976, + 0.0026953655760735273, + 0.0021016814280301332, + 0.0015901861479505897, + 0.0012815530644729733, + 0.0008121848222799599, + 0.0009081335156224668, + 0.0013344736071303487, + 0.0011041766265407205, + 0.001423766021616757, + 0.00189974473323673, + 0.0023410264402627945, + 0.0025312029756605625, + 0.0016024709912016988, + 0.002409544540569186, + 0.0014263753546401858, + 0.0022795326076447964, + 0.001934211584739387, + 0.001690707285888493, + 0.005823508836328983, + 0.0012122277403250337, + 0.0009104397031478584, + 0.0014622361632063985, + 0.0011230935342609882, + 0.0009597907192073762, + 0.0032835875172168016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.005948459263890982 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370697423815727 + ], + "opposite_logprob_tokens": [ + -3.4087069034576416 + ], + "logprob_reward": -0.03370697423815727, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370697423815727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0617382787168026, + 0.8052961230278015, + 0.012114355340600014, + 0.013037529774010181, + 0.02204469032585621, + 0.012015257962048054, + 0.03859773650765419, + 0.011840916238725185, + 0.010699317790567875, + 0.012615843676030636 + ], + "credit_map": [ + 0.00711611844599247, + 0.003283414291217923, + 0.003773827338591218, + 0.0017970561748370528, + 0.005505325738340616, + 0.016069775447249413, + 0.0046260119415819645, + 0.019566746428608894, + 0.11125633865594864, + 0.04842269420623779, + 0.03339853882789612, + 0.024122821167111397, + 0.09028834849596024, + 0.1822834461927414, + 0.043123409152030945, + 0.272400438785553, + 0.0016595234628766775, + 0.0018034273525699973, + 0.0010384973138570786, + 0.0009761677356436849, + 0.0025035652797669172, + 0.001758137485012412, + 0.0008617622079327703, + 0.0015132729895412922, + 0.0012614992447197437, + 0.0016716732643544674, + 0.0009105888311751187, + 0.0009251430747099221, + 0.0034243552945554256, + 0.0023738399613648653, + 0.0010456385789439082, + 0.0014247903600335121, + 0.0028466051444411278, + 0.0019244354916736484, + 0.0021151104010641575, + 0.002363341860473156, + 0.004282375797629356, + 0.0032064432743936777, + 0.002331480383872986, + 0.002974897623062134, + 0.0015440095448866487, + 0.0013389030937105417, + 0.0013120434014126658, + 0.0014005765551701188, + 0.002330027287825942, + 0.0020250482484698296, + 0.001151443226262927, + 0.0009132047998718917, + 0.004203364253044128, + 0.00354363233782351, + 0.0022301790304481983, + 0.0017940058605745435, + 0.012203410267829895, + 0.00825712364166975, + 0.002462344476953149, + 0.003903674893081188, + 0.001594700152054429, + 0.002159366849809885, + 0.0016827400540933013, + 0.0011993972584605217, + 0.001119320746511221, + 0.0012924823677167296, + 0.0013343517202883959, + 0.00145855569280684, + 0.0015196208842098713, + 0.0015474880347028375, + 0.001336017856374383, + 0.0006530066020786762, + 0.0018248248379677534, + 0.0013173471670597792, + 0.001235230010934174, + 0.0012657821644097567, + 0.0014799939235672355, + 0.003466001944616437, + 0.0010222382843494415, + 0.0007210313342511654, + 0.0015179621987044811, + 0.001177322119474411, + 0.001034770393744111, + 0.002196522895246744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.03370697423815727 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704721093177795 + ], + "opposite_logprob_tokens": [ + -2.127047300338745 + ], + "logprob_reward": -0.12704721093177795, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704721093177795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11302517354488373, + 0.7586922645568848, + 0.06984757632017136, + 0.008470156230032444, + 0.008001605980098248, + 0.009154075756669044, + 0.014155146665871143, + 0.007926256395876408, + 0.00427560368552804, + 0.006452107336372137 + ], + "credit_map": [ + 0.013457526452839375, + 0.004924265667796135, + 0.0044930800795555115, + 0.0019482729258015752, + 0.01098821684718132, + 0.03310147300362587, + 0.005036973860114813, + 0.0390753448009491, + 0.1048625037074089, + 0.04494917020201683, + 0.03094143606722355, + 0.02240709215402603, + 0.08605409413576126, + 0.170676127076149, + 0.040856458246707916, + 0.25794532895088196, + 0.008682504296302795, + 0.004514922853559256, + 0.003038658294826746, + 0.002812279388308525, + 0.00816167239099741, + 0.01751374453306198, + 0.003842840902507305, + 0.021280944347381592, + 0.000889404327608645, + 0.0013055992312729359, + 0.0007089825812727213, + 0.0008937394013628364, + 0.0017507156589999795, + 0.0013335042167454958, + 0.0007737865089438856, + 0.0008144237799569964, + 0.0008792108856141567, + 0.0008517917012795806, + 0.000651540351100266, + 0.0008424505940638483, + 0.0013000498292967677, + 0.0012531005777418613, + 0.0008674863493070006, + 0.001355974469333887, + 0.0008161335717886686, + 0.0008354135788977146, + 0.0006286626448854804, + 0.0006146630621515214, + 0.0019250042969360948, + 0.0018362243426963687, + 0.0008102149586193264, + 0.0016877584857866168, + 0.001249591470696032, + 0.001164457411505282, + 0.0010735506657510996, + 0.0007666504243388772, + 0.004896556492894888, + 0.0025381145533174276, + 0.0008806133992038667, + 0.0015856105601415038, + 0.001164023531600833, + 0.0011434347834438086, + 0.0010911936406046152, + 0.0007033705478534102, + 0.0007818798767402768, + 0.0010478277690708637, + 0.0007612496265210211, + 0.0012332753976806998, + 0.0006739970995113254, + 0.0005224276683293283, + 0.0006247470737434924, + 0.0003707412979565561, + 0.0006002490408718586, + 0.0005035203066654503, + 0.0004429072141647339, + 0.0005370135186240077, + 0.000561200431548059, + 0.002397448057308793, + 0.00029180647106841207, + 0.0002756138565018773, + 0.000510423444211483, + 0.00047737901331856847, + 0.00030489038908854127, + 0.0016333444509655237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.12704721093177795 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An eye-level close-up, capturing the subtle nervous tremor of a handheld camera, reveals a massive great white shark suspended, motionless, in the vast abyss. The shark's massive, scarred body, a gradient of dark slate gray on top and a distinctly white underbelly, fills the frame, its skin textured with the subtle roughness of dense dermis. Surrounded by an endless expanse of deep cobalt blue, faint, ethereal sunlight penetrates the water from above, catching flecks of marine snow and nearly illuminating its dorsal flank before being swallowed by the overwhelming darkness below. Its body is a masterpiece of predatory design, held in eerie stillness, poised for an imminent, explosive attack. Its left eye, a black, unblinking, soulless orb, is fixed directly on the lens, creating an unsettling and direct confrontation, while its jaw, slightly parted, reveals rows of jagged, triangular teeth. The only perceptible movement is a slight, eerie tremor of the camera and the slow, rhythmic opening of the shark's gills, creating a chilling atmosphere of pure, breathless tension and primal fear.", + "seed": 3933, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_4/index_3628_seed_3933_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.052847223449498416, + "vlm_logprob": -0.052847223449498416, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.052847223449498416, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054047707468271255 + ], + "opposite_logprob_tokens": [ + -3.3040478229522705 + ], + "logprob_reward": -0.054047707468271255, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.054047707468271255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14534777402877808, + 0.17092040181159973, + 0.07620635628700256, + 0.14259269833564758, + 0.0808349996805191, + 0.1701377034187317, + 0.08738885074853897, + 0.03449936583638191, + 0.047297149896621704, + 0.04477463290095329 + ], + "credit_map": [ + 0.019477959722280502, + 0.014584386721253395, + 0.018886107951402664, + 0.01737806759774685, + 0.017673486843705177, + 0.02090420201420784, + 0.016536854207515717, + 0.019906703382730484, + 0.018693452700972557, + 0.023700160905718803, + 0.025233980268239975, + 0.020385773852467537, + 0.02257418818771839, + 0.02308533526957035, + 0.016008025035262108, + 0.021239491179585457, + 0.006821696646511555, + 0.008332058787345886, + 0.010613802820444107, + 0.006387598346918821, + 0.012107189744710922, + 0.010740778408944607, + 0.00957136694341898, + 0.011631866917014122, + 0.011970899999141693, + 0.021096497774124146, + 0.02039538323879242, + 0.01542192418128252, + 0.016092674806714058, + 0.020245281979441643, + 0.013986767269670963, + 0.02338327094912529, + 0.009736432693898678, + 0.0085003562271595, + 0.014642603695392609, + 0.008083779364824295, + 0.008117645978927612, + 0.009644352830946445, + 0.00706326263025403, + 0.015046568587422371, + 0.020707860589027405, + 0.014984835870563984, + 0.043144699186086655, + 0.01323268935084343, + 0.018217597156763077, + 0.019550208002328873, + 0.010042202658951283, + 0.03025761805474758, + 0.009407318197190762, + 0.008085737004876137, + 0.018587930127978325, + 0.0074686696752905846, + 0.013321940787136555, + 0.015344503335654736, + 0.0075707947835326195, + 0.007601955905556679, + 0.003210145980119705, + 0.0038088306318968534, + 0.005051424726843834, + 0.0026828027330338955, + 0.004864098038524389, + 0.004226955585181713, + 0.005603698082268238, + 0.005051411222666502, + 0.007045050151646137, + 0.006052033510059118, + 0.007531530223786831, + 0.0038998029194772243, + 0.005560220684856176, + 0.00578722357749939, + 0.005750051233917475, + 0.005671234335750341, + 0.0070512425154447556, + 0.004978924989700317, + 0.005605833139270544, + 0.003370647318661213, + 0.007119290065020323, + 0.005720362067222595, + 0.005360386334359646, + 0.005567946936935186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.054047707468271255 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011099381372332573 + ], + "opposite_logprob_tokens": [ + -4.511099338531494 + ], + "logprob_reward": -0.011099381372332573, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011099381372332573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1979263871908188, + 0.06731123477220535, + 0.15815313160419464, + 0.15164989233016968, + 0.0837085172533989, + 0.05114623159170151, + 0.06548324227333069, + 0.09656458348035812, + 0.06741146743297577, + 0.060645341873168945 + ], + "credit_map": [ + 0.022167470306158066, + 0.02085561491549015, + 0.02450842596590519, + 0.02079669199883938, + 0.028942758217453957, + 0.03215121477842331, + 0.02342929318547249, + 0.025074904784560204, + 0.0069388458505272865, + 0.007187890820205212, + 0.008952701464295387, + 0.008023454807698727, + 0.008419842459261417, + 0.009212320670485497, + 0.008716276846826077, + 0.009859895333647728, + 0.013719036243855953, + 0.020137419924139977, + 0.020396269857883453, + 0.018157755956053734, + 0.021106209605932236, + 0.025847570970654488, + 0.016878766939044, + 0.021910106763243675, + 0.013871137984097004, + 0.01928640343248844, + 0.023762298747897148, + 0.01207746472209692, + 0.015020537190139294, + 0.020749961957335472, + 0.015973659232258797, + 0.03090842068195343, + 0.008376263082027435, + 0.005431399680674076, + 0.01809692569077015, + 0.005913072265684605, + 0.007029403932392597, + 0.009576461277902126, + 0.01036741305142641, + 0.01891757734119892, + 0.007454775273799896, + 0.00589399691671133, + 0.00892206933349371, + 0.003914389759302139, + 0.006448308005928993, + 0.004611124284565449, + 0.0049687959253787994, + 0.008932771161198616, + 0.006173956673592329, + 0.006929779890924692, + 0.015334031544625759, + 0.004595702514052391, + 0.008137189783155918, + 0.005272905807942152, + 0.005182604305446148, + 0.013857068493962288, + 0.008055693469941616, + 0.010124631226062775, + 0.015489255078136921, + 0.007529319729655981, + 0.013372228480875492, + 0.010556385852396488, + 0.02064601518213749, + 0.010791047476232052, + 0.008240853436291218, + 0.0068619102239608765, + 0.01632104255259037, + 0.005199607461690903, + 0.00654767919331789, + 0.006294977851212025, + 0.006388578098267317, + 0.011556815356016159, + 0.010605573654174805, + 0.008475638926029205, + 0.008588130585849285, + 0.006038422230631113, + 0.007599266245961189, + 0.004620389547199011, + 0.008478044532239437, + 0.006239871494472027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.011099381372332573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631078828126192 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631078828126192, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631078828126192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10799525678157806, + 0.07753569632768631, + 0.12008566409349442, + 0.0529337003827095, + 0.28046485781669617, + 0.04063112288713455, + 0.10230189561843872, + 0.13568025827407837, + 0.03246086835861206, + 0.049910590052604675 + ], + "credit_map": [ + 0.01878507249057293, + 0.012941022403538227, + 0.018128883093595505, + 0.008583184331655502, + 0.015614999458193779, + 0.011579482816159725, + 0.008289541117846966, + 0.014073076657950878, + 0.01392186339944601, + 0.00634244829416275, + 0.01354339811950922, + 0.007893508300185204, + 0.006031535565853119, + 0.007901744917035103, + 0.007082635071128607, + 0.014818566851317883, + 0.012126777321100235, + 0.012058745138347149, + 0.022398345172405243, + 0.010999061167240143, + 0.009515843354165554, + 0.0120807234197855, + 0.015560531057417393, + 0.025345642119646072, + 0.005330292508006096, + 0.0074718790128827095, + 0.007986120879650116, + 0.005269401241093874, + 0.007883126847445965, + 0.005515638273209333, + 0.004667195957154036, + 0.00881004985421896, + 0.034041255712509155, + 0.024034220725297928, + 0.04392237216234207, + 0.02076665312051773, + 0.024494633078575134, + 0.042215973138809204, + 0.02308337390422821, + 0.06790640205144882, + 0.004473565611988306, + 0.005068240687251091, + 0.0061960285529494286, + 0.0031669382005929947, + 0.005622574128210545, + 0.00598288094624877, + 0.004019679501652718, + 0.006101218983530998, + 0.01015785988420248, + 0.00893267709761858, + 0.02953959070146084, + 0.005557771772146225, + 0.009266404435038567, + 0.023570429533720016, + 0.006025707349181175, + 0.009251459501683712, + 0.015952518209815025, + 0.015450669452548027, + 0.021388988941907883, + 0.010477535426616669, + 0.017180832102894783, + 0.013329633511602879, + 0.024025581777095795, + 0.01787450909614563, + 0.0037327371537685394, + 0.0035382977221161127, + 0.004764187149703503, + 0.0025290213525295258, + 0.0029952405020594597, + 0.0037100433837622404, + 0.004684717394411564, + 0.006506627891212702, + 0.008180785924196243, + 0.0069399732165038586, + 0.006498632486909628, + 0.004216399975121021, + 0.0059853107668459415, + 0.005590368062257767, + 0.006736963987350464, + 0.005762164946645498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.007631078828126192 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865852743387222 + ], + "opposite_logprob_tokens": [ + -3.048658609390259 + ], + "logprob_reward": -0.04865852743387222, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865852743387222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2236213982105255, + 0.06599846482276917, + 0.2400614470243454, + 0.08383115381002426, + 0.09569277614355087, + 0.04423576965928078, + 0.042842213064432144, + 0.05314259231090546, + 0.04041348770260811, + 0.1101607084274292 + ], + "credit_map": [ + 0.025014536455273628, + 0.023211000487208366, + 0.02768317610025406, + 0.019521865993738174, + 0.03941715508699417, + 0.03727606311440468, + 0.023198232054710388, + 0.02829936146736145, + 0.007068889681249857, + 0.008931715972721577, + 0.008596747182309628, + 0.004832983482629061, + 0.011419695802032948, + 0.0079543087631464, + 0.0066716838628053665, + 0.01052243821322918, + 0.021546179428696632, + 0.03371556103229523, + 0.03797212243080139, + 0.018330316990613937, + 0.030982986092567444, + 0.033643417060375214, + 0.018840983510017395, + 0.04502987489104271, + 0.008221848867833614, + 0.01397805754095316, + 0.01483985036611557, + 0.007720393594354391, + 0.007945210672914982, + 0.011984182521700859, + 0.006599467247724533, + 0.01254214160144329, + 0.00960161630064249, + 0.010288751684129238, + 0.016344206407666206, + 0.007878171280026436, + 0.008837486617267132, + 0.014547498896718025, + 0.006671493407338858, + 0.021523546427488327, + 0.0057782502844929695, + 0.005022278521209955, + 0.009665366262197495, + 0.003056233050301671, + 0.005405441392213106, + 0.005280535668134689, + 0.003389058168977499, + 0.006638605613261461, + 0.004146887455135584, + 0.004347796086221933, + 0.008741219528019428, + 0.0032940558157861233, + 0.0057100560516119, + 0.007836262695491314, + 0.003741147229447961, + 0.005024788435548544, + 0.0043282764963805676, + 0.004459712188690901, + 0.011219041422009468, + 0.0036067303735762835, + 0.007318928372114897, + 0.005366665776818991, + 0.008858654648065567, + 0.007984583266079426, + 0.006116053089499474, + 0.004843675997108221, + 0.0055217077024281025, + 0.0030202826019376516, + 0.004956922028213739, + 0.004575812257826328, + 0.006792690604925156, + 0.00458634365350008, + 0.02045164629817009, + 0.01327100582420826, + 0.01725844293832779, + 0.00705856503918767, + 0.015778686851263046, + 0.009464449249207973, + 0.012345405295491219, + 0.014532506465911865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.04865852743387222 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279942214488983 + ], + "opposite_logprob_tokens": [ + -2.0177993774414062 + ], + "logprob_reward": -0.14279942214488983, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279942214488983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14363932609558105, + 0.07390299439430237, + 0.12936289608478546, + 0.11757981032133102, + 0.09548906236886978, + 0.10630811005830765, + 0.1220650002360344, + 0.08049598336219788, + 0.04629833623766899, + 0.0848584696650505 + ], + "credit_map": [ + 0.014850731939077377, + 0.014675246551632881, + 0.01758430339396, + 0.012113924138247967, + 0.025740154087543488, + 0.02073417790234089, + 0.014675355516374111, + 0.023265432566404343, + 0.007374722510576248, + 0.009156008251011372, + 0.01000208780169487, + 0.008802596479654312, + 0.009174668230116367, + 0.008391139097511768, + 0.0079151950776577, + 0.013086576946079731, + 0.011008466593921185, + 0.013171102851629257, + 0.021601414307951927, + 0.01595952734351158, + 0.011341958306729794, + 0.013479827903211117, + 0.024687711149454117, + 0.018112890422344208, + 0.0179181806743145, + 0.013634644448757172, + 0.017010489478707314, + 0.008520851843059063, + 0.011288165114820004, + 0.01757723093032837, + 0.010211586020886898, + 0.021418657153844833, + 0.012690149247646332, + 0.008379009552299976, + 0.021305490285158157, + 0.007430234458297491, + 0.009676340036094189, + 0.011489946395158768, + 0.008487033657729626, + 0.01603085547685623, + 0.015777816995978355, + 0.016543574631214142, + 0.012861225754022598, + 0.00567628163844347, + 0.017763718962669373, + 0.014750233851373196, + 0.00947569590061903, + 0.01345956139266491, + 0.011458516120910645, + 0.009030732326209545, + 0.031491752713918686, + 0.00717496033757925, + 0.014645049348473549, + 0.023846590891480446, + 0.010671056807041168, + 0.013746339827775955, + 0.006631845608353615, + 0.01142565906047821, + 0.012725216336548328, + 0.005509695969521999, + 0.013910987414419651, + 0.010208976455032825, + 0.00844282191246748, + 0.011640786193311214, + 0.006764429621398449, + 0.005370864178985357, + 0.006984381936490536, + 0.0034634671173989773, + 0.0049535250291228294, + 0.0061205290257930756, + 0.005852866917848587, + 0.00678827241063118, + 0.006066453643143177, + 0.008831631392240524, + 0.01229243353009224, + 0.005852559581398964, + 0.012513220310211182, + 0.011379220522940159, + 0.015553340315818787, + 0.012369615957140923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.14279942214488983 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An eye-level close-up, capturing the subtle nervous tremor of a handheld camera, reveals a massive great white shark suspended, motionless, in the vast abyss. The shark's massive, scarred body, a gradient of dark slate gray on top and a distinctly white underbelly, fills the frame, its skin textured with the subtle roughness of dense dermis. Surrounded by an endless expanse of deep cobalt blue, faint, ethereal sunlight penetrates the water from above, catching flecks of marine snow and nearly illuminating its dorsal flank before being swallowed by the overwhelming darkness below. Its body is a masterpiece of predatory design, held in eerie stillness, poised for an imminent, explosive attack. Its left eye, a black, unblinking, soulless orb, is fixed directly on the lens, creating an unsettling and direct confrontation, while its jaw, slightly parted, reveals rows of jagged, triangular teeth. The only perceptible movement is a slight, eerie tremor of the camera and the slow, rhythmic opening of the shark's gills, creating a chilling atmosphere of pure, breathless tension and primal fear.", + "seed": 3934, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_4/index_3628_seed_3934_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07219206858426333, + "vlm_logprob": -0.07219206858426333, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.07219206858426333, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0779956802725792 + ], + "opposite_logprob_tokens": [ + -2.952995777130127 + ], + "logprob_reward": -0.0779956802725792, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0779956802725792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42531439661979675, + 0.024873359128832817, + 0.0590212345123291, + 0.05046670511364937, + 0.04447515681385994, + 0.04866649582982063, + 0.16690611839294434, + 0.024168003350496292, + 0.12352398782968521, + 0.03258449211716652 + ], + "credit_map": [ + 0.040976785123348236, + 0.010469096712768078, + 0.11433454602956772, + 0.011274386197328568, + 0.023698262870311737, + 0.15861178934574127, + 0.017487604171037674, + 0.0484619066119194, + 0.001923011033795774, + 0.0009286123095080256, + 0.006563378032296896, + 0.0012936118291690946, + 0.002357702236622572, + 0.006443768739700317, + 0.00241525168530643, + 0.0029480240773409605, + 0.004083793610334396, + 0.0015353697817772627, + 0.0216006301343441, + 0.0023207012563943863, + 0.002571013756096363, + 0.021829767152667046, + 0.0017663674661889672, + 0.0033135937992483377, + 0.004794017411768436, + 0.003292249282822013, + 0.008396041579544544, + 0.0043428922072052956, + 0.00701146712526679, + 0.00831436738371849, + 0.00489352410659194, + 0.00942214671522379, + 0.003601633943617344, + 0.003367368131875992, + 0.009553434327244759, + 0.004543588031083345, + 0.005254535470157862, + 0.008026709780097008, + 0.0033732966985553503, + 0.006754590664058924, + 0.004641735460609198, + 0.0028911170084029436, + 0.009400850161910057, + 0.0035978243686258793, + 0.006535291206091642, + 0.008151923306286335, + 0.005907955579459667, + 0.00753979803994298, + 0.006754429079592228, + 0.002609127201139927, + 0.06242646649479866, + 0.004329303745180368, + 0.012156053446233273, + 0.07119384407997131, + 0.0017198837595060468, + 0.00571700744330883, + 0.0020890175364911556, + 0.002497978275641799, + 0.0035163876600563526, + 0.001987988129258156, + 0.003714917693287134, + 0.003289076965302229, + 0.0030356680508702993, + 0.004036969505250454, + 0.009494613856077194, + 0.010739732533693314, + 0.017679866403341293, + 0.005387491080909967, + 0.040884871035814285, + 0.021316293627023697, + 0.010255281813442707, + 0.007765840273350477, + 0.003910160157829523, + 0.00312585080973804, + 0.005130202043801546, + 0.001631758757866919, + 0.006096395198255777, + 0.005452408920973539, + 0.0027204789221286774, + 0.004517237655818462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.0779956802725792 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012567962519824505 + ], + "opposite_logprob_tokens": [ + -4.38756799697876 + ], + "logprob_reward": -0.012567962519824505, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012567962519824505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056403014808893204, + 0.1612364649772644, + 0.027747249230742455, + 0.012266288511455059, + 0.020344240590929985, + 0.008390807546675205, + 0.4549444317817688, + 0.01343468576669693, + 0.21814864873886108, + 0.02708417735993862 + ], + "credit_map": [ + 0.004836356267333031, + 0.0021584811620414257, + 0.015001010149717331, + 0.0033178699668496847, + 0.004354078322649002, + 0.017827220261096954, + 0.003779950551688671, + 0.005128050688654184, + 0.009692652150988579, + 0.002714752219617367, + 0.0634850487112999, + 0.005468375980854034, + 0.004619513172656298, + 0.06597843766212463, + 0.003698840970173478, + 0.005578845273703337, + 0.0032606476452201605, + 0.0018497962737455964, + 0.0031333251390606165, + 0.0017447499558329582, + 0.0053879264742136, + 0.0038321558386087418, + 0.0032864680979400873, + 0.005252183880656958, + 0.0010398086160421371, + 0.0008259088499471545, + 0.0023105430882424116, + 0.0010602765250951052, + 0.0015453505329787731, + 0.0023287213407456875, + 0.0013558409409597516, + 0.0017998385010287166, + 0.001608330407179892, + 0.001594369183294475, + 0.0030841375701129436, + 0.0017653985414654016, + 0.002933682408183813, + 0.002554767532274127, + 0.002013235120102763, + 0.004790318198502064, + 0.0006080979364924133, + 0.0004538540088105947, + 0.0020828612614423037, + 0.0004933194140903652, + 0.0008929198374971747, + 0.0021741383243352175, + 0.0007492589647881687, + 0.0009363575954921544, + 0.018243832513689995, + 0.0068550072610378265, + 0.17000456154346466, + 0.011641873978078365, + 0.03382546082139015, + 0.1942172348499298, + 0.004609616473317146, + 0.015546840615570545, + 0.0011100505944341421, + 0.0011938967509195209, + 0.0017577678663656116, + 0.001224170089699328, + 0.001777594443410635, + 0.0016688489122316241, + 0.002423287834972143, + 0.002279069507494569, + 0.01658557914197445, + 0.018644830211997032, + 0.03065420128405094, + 0.009359092451632023, + 0.07225692272186279, + 0.03839927539229393, + 0.018127543851733208, + 0.01412120833992958, + 0.0027739403303712606, + 0.003110096324235201, + 0.004342131782323122, + 0.002097716787829995, + 0.004246192052960396, + 0.003842039732262492, + 0.003908940590918064, + 0.002763119526207447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.012567962519824505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741398945450783 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741398945450783, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741398945450783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41567063331604004, + 0.12175244837999344, + 0.06153343245387077, + 0.016396397724747658, + 0.029155096039175987, + 0.011164720170199871, + 0.29038047790527344, + 0.03507479280233383, + 0.007870701141655445, + 0.011001357808709145 + ], + "credit_map": [ + 0.03765879198908806, + 0.010528000071644783, + 0.11321932077407837, + 0.011602138169109821, + 0.0243221502751112, + 0.15429429709911346, + 0.017666032537817955, + 0.04637986421585083, + 0.007282000966370106, + 0.001664222334511578, + 0.047781143337488174, + 0.003905956633388996, + 0.0036571931559592485, + 0.0502881221473217, + 0.0026555468793958426, + 0.004518247675150633, + 0.00407281331717968, + 0.0016194360796362162, + 0.02159639075398445, + 0.003324450459331274, + 0.0027891595382243395, + 0.02135472744703293, + 0.0027820062823593616, + 0.0039944397285580635, + 0.0019269014010205865, + 0.0009361639968119562, + 0.00330307031981647, + 0.0013459945330396295, + 0.0020356362219899893, + 0.002381640486419201, + 0.001473357086069882, + 0.0029936337377876043, + 0.002856591483578086, + 0.001871013082563877, + 0.00448860228061676, + 0.002553149126470089, + 0.00412799371406436, + 0.003376139560714364, + 0.0029862469527870417, + 0.0068953558802604675, + 0.0008562355651520193, + 0.0007574124028906226, + 0.0028671417385339737, + 0.0007179781678132713, + 0.0010577960638329387, + 0.0030393649358302355, + 0.0007663512951694429, + 0.001102438778616488, + 0.01213522907346487, + 0.004782660398632288, + 0.1077302098274231, + 0.007836836390197277, + 0.021657783538103104, + 0.12297331541776657, + 0.003204392734915018, + 0.010060016065835953, + 0.003049144521355629, + 0.0038594587240368128, + 0.0051284716464579105, + 0.0034239734522998333, + 0.005216746125370264, + 0.004560042172670364, + 0.004283936228603125, + 0.0055530178360641, + 0.000705020735040307, + 0.0007919750059954822, + 0.0013521220535039902, + 0.00045564695028588176, + 0.0014616706175729632, + 0.0010231096530333161, + 0.0008848966681398451, + 0.0011962588177993894, + 0.0009214082383550704, + 0.0008435052004642785, + 0.0022127872798591852, + 0.0007790671661496162, + 0.0013930790591984987, + 0.002795218490064144, + 0.0011282616760581732, + 0.0009280295926146209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.006741398945450783 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062068115919828415 + ], + "opposite_logprob_tokens": [ + -2.812068223953247 + ], + "logprob_reward": -0.062068115919828415, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062068115919828415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5419136881828308, + 0.04852733016014099, + 0.015464268624782562, + 0.01886807382106781, + 0.018825558945536613, + 0.014455510303378105, + 0.1615738719701767, + 0.008497598581016064, + 0.15460866689682007, + 0.017265407368540764 + ], + "credit_map": [ + 0.05117517337203026, + 0.013099420815706253, + 0.14754147827625275, + 0.014537526294589043, + 0.03052799589931965, + 0.2023521214723587, + 0.0213407464325428, + 0.06133922189474106, + 0.0029657785780727863, + 0.0012348712189123034, + 0.01688631996512413, + 0.0016809793887659907, + 0.0027660278137773275, + 0.018226217478513718, + 0.002203157404437661, + 0.0025639794766902924, + 0.0012724451953545213, + 0.0005848290165886283, + 0.004296587314456701, + 0.0008875697385519743, + 0.0013018813915550709, + 0.004482395481318235, + 0.0010987017303705215, + 0.0015398601535707712, + 0.0012335567735135555, + 0.0010288608027622104, + 0.004133273381739855, + 0.0013859362807124853, + 0.0020204302854835987, + 0.0049253362230956554, + 0.00123310717754066, + 0.0029075732454657555, + 0.0017877843929454684, + 0.0014188602799549699, + 0.0035892680753022432, + 0.00172201928216964, + 0.002237883862107992, + 0.0031828666105866432, + 0.0015785607974976301, + 0.00330831715837121, + 0.0012278174981474876, + 0.0008284474024549127, + 0.0027029092889279127, + 0.001158274826593697, + 0.001885006669908762, + 0.0028469490353018045, + 0.0015542798209935427, + 0.002251825761049986, + 0.006870179437100887, + 0.003089760895818472, + 0.05894911289215088, + 0.004799174144864082, + 0.011812953278422356, + 0.06724661588668823, + 0.0021186594385653734, + 0.006687410641461611, + 0.0010704725282266736, + 0.0007892688736319542, + 0.0011084398720413446, + 0.000534428283572197, + 0.0016128694405779243, + 0.0011814199388027191, + 0.0010834637796506286, + 0.0011172355152666569, + 0.011842596344649792, + 0.0134927062317729, + 0.02159682847559452, + 0.006614586338400841, + 0.051663629710674286, + 0.027124915271997452, + 0.01271284930408001, + 0.009560558013617992, + 0.001910834456793964, + 0.0018956781132146716, + 0.0035830691922456026, + 0.0008628563373349607, + 0.002227175049483776, + 0.0031799653079360723, + 0.0014681856846436858, + 0.0021376425866037607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.062068115919828415 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158718526363373 + ], + "opposite_logprob_tokens": [ + -1.701587200164795 + ], + "logprob_reward": -0.20158718526363373, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158718526363373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7556150555610657, + 0.05200878158211708, + 0.07108592987060547, + 0.00899007823318243, + 0.0155467689037323, + 0.019227394834160805, + 0.018081199377775192, + 0.004791132640093565, + 0.04997187852859497, + 0.004681767895817757 + ], + "credit_map": [ + 0.06910909712314606, + 0.01806705631315708, + 0.20778286457061768, + 0.020278779789805412, + 0.04273856431245804, + 0.2825692296028137, + 0.030612312257289886, + 0.08445717394351959, + 0.0030507505871355534, + 0.0009247040143236518, + 0.02043050527572632, + 0.001575086498633027, + 0.0024665568489581347, + 0.020015103742480278, + 0.0012879119021818042, + 0.002258160151541233, + 0.0056452383287250996, + 0.0014665882335975766, + 0.027164066210389137, + 0.0026703490875661373, + 0.00324981939047575, + 0.026108013466000557, + 0.001893768785521388, + 0.002888085786253214, + 0.0008782665245234966, + 0.00040309186442755163, + 0.0021647124085575342, + 0.0005826594424434006, + 0.0006461673765443265, + 0.0026447600685060024, + 0.0005336512695066631, + 0.0011367698898538947, + 0.001520938822068274, + 0.001384696108289063, + 0.003698958083987236, + 0.0012344693532213569, + 0.0018049123464152217, + 0.0031277528032660484, + 0.0010460811899974942, + 0.0017289597308263183, + 0.0016358650755137205, + 0.0007952446467243135, + 0.005238366313278675, + 0.0008621285087428987, + 0.0019199175294488668, + 0.005978813860565424, + 0.0011975517263635993, + 0.001599508454091847, + 0.0016462156781926751, + 0.0012193636503070593, + 0.004405390936881304, + 0.0010239267721772194, + 0.001631160848774016, + 0.005286573898047209, + 0.000806300260592252, + 0.0020622669253498316, + 0.00040818023262545466, + 0.0004622544511221349, + 0.0006369654438458383, + 0.00038421741919592023, + 0.0009096887079067528, + 0.0007148470031097531, + 0.000625454995315522, + 0.000649524328764528, + 0.003839705605059862, + 0.004334847908467054, + 0.0068162595853209496, + 0.0021148896776139736, + 0.016611233353614807, + 0.008850619196891785, + 0.004196935798972845, + 0.0032073864713311195, + 0.0003869391803164035, + 0.0003562348720151931, + 0.000818870379589498, + 0.00027006707387045026, + 0.0006929222727194428, + 0.0011479532113298774, + 0.000592279015108943, + 0.0004165016289334744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.20158718526363373 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An eye-level close-up, capturing the subtle nervous tremor of a handheld camera, reveals a massive great white shark suspended, motionless, in the vast abyss. The shark's massive, scarred body, a gradient of dark slate gray on top and a distinctly white underbelly, fills the frame, its skin textured with the subtle roughness of dense dermis. Surrounded by an endless expanse of deep cobalt blue, faint, ethereal sunlight penetrates the water from above, catching flecks of marine snow and nearly illuminating its dorsal flank before being swallowed by the overwhelming darkness below. Its body is a masterpiece of predatory design, held in eerie stillness, poised for an imminent, explosive attack. Its left eye, a black, unblinking, soulless orb, is fixed directly on the lens, creating an unsettling and direct confrontation, while its jaw, slightly parted, reveals rows of jagged, triangular teeth. The only perceptible movement is a slight, eerie tremor of the camera and the slow, rhythmic opening of the shark's gills, creating a chilling atmosphere of pure, breathless tension and primal fear.", + "seed": 3935, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_4/index_3628_seed_3935_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05886142086237669, + "vlm_logprob": -0.05886142086237669, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.05886142086237669, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11258804053068161 + ], + "opposite_logprob_tokens": [ + -2.3625879287719727 + ], + "logprob_reward": -0.11258804053068161, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11258804053068161, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06573795527219772, + 0.061265505850315094, + 0.03156334161758423, + 0.05305143818259239, + 0.014737807214260101, + 0.023190010339021683, + 0.6154426336288452, + 0.06692492216825485, + 0.03681781142950058, + 0.03126855939626694 + ], + "credit_map": [ + 0.012860649265348911, + 0.00903191976249218, + 0.007798300124704838, + 0.005351333413273096, + 0.006154257338494062, + 0.010156148113310337, + 0.005866282153874636, + 0.008519061841070652, + 0.004411362111568451, + 0.0033625192008912563, + 0.0047044409438967705, + 0.003862231969833374, + 0.004339240957051516, + 0.01614147238433361, + 0.007586462888866663, + 0.016857773065567017, + 0.008788669481873512, + 0.0032117075752466917, + 0.0030887245666235685, + 0.0017242026515305042, + 0.0020398495253175497, + 0.005231277551501989, + 0.0019766127225011587, + 0.005502295680344105, + 0.005871871020644903, + 0.005830478388816118, + 0.006144554819911718, + 0.00782363023608923, + 0.006348612252622843, + 0.0069427150301635265, + 0.00821303017437458, + 0.0058765411376953125, + 0.001799721852876246, + 0.0019948433618992567, + 0.0017829565331339836, + 0.0019141845405101776, + 0.0019371678354218602, + 0.001956535968929529, + 0.0016965330578386784, + 0.001655863830819726, + 0.0031090248376131058, + 0.004605039954185486, + 0.002488178201019764, + 0.0023473852779716253, + 0.002690868452191353, + 0.0032731955870985985, + 0.002128652762621641, + 0.0025476652663201094, + 0.21371902525424957, + 0.022585924714803696, + 0.02749151550233364, + 0.01241874136030674, + 0.04261549189686775, + 0.2533072233200073, + 0.0117937121540308, + 0.031511008739471436, + 0.010793721303343773, + 0.006324256304651499, + 0.011129496619105339, + 0.0068521322682499886, + 0.005927050486207008, + 0.011354604735970497, + 0.003629176178947091, + 0.010914486832916737, + 0.005741502624005079, + 0.005480777472257614, + 0.006175405345857143, + 0.004787845071405172, + 0.0033994740806519985, + 0.004073186777532101, + 0.0025985827669501305, + 0.00456103915348649, + 0.005471827462315559, + 0.005399114452302456, + 0.004360420163720846, + 0.0031353808008134365, + 0.003075459273532033, + 0.0028432528488337994, + 0.0025346039328724146, + 0.004448499530553818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.11258804053068161 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421845331788063 + ], + "opposite_logprob_tokens": [ + -4.264218330383301 + ], + "logprob_reward": -0.01421845331788063, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421845331788063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16840045154094696, + 0.08509357273578644, + 0.05975392460823059, + 0.05083887651562691, + 0.04767784848809242, + 0.2583782374858856, + 0.15122540295124054, + 0.10351672023534775, + 0.0413813516497612, + 0.03373360633850098 + ], + "credit_map": [ + 0.02291046641767025, + 0.01343960128724575, + 0.00922953337430954, + 0.010671922005712986, + 0.011869065463542938, + 0.04260854423046112, + 0.01124642789363861, + 0.04642489552497864, + 0.006601286586374044, + 0.004546418320387602, + 0.006182173732668161, + 0.004124372266232967, + 0.004573405254632235, + 0.02577325329184532, + 0.0064615304581820965, + 0.026831129565835, + 0.010126734152436256, + 0.007149262819439173, + 0.007602953352034092, + 0.004262868780642748, + 0.00875876098871231, + 0.007510887458920479, + 0.007037778850644827, + 0.007304678205400705, + 0.006426418200135231, + 0.004203793592751026, + 0.007438282947987318, + 0.005213878583163023, + 0.005729496479034424, + 0.007484531961381435, + 0.008812825195491314, + 0.005529649555683136, + 0.004082071594893932, + 0.006306900177150965, + 0.0071622394025325775, + 0.0072921449318528175, + 0.0076180375181138515, + 0.006543517578393221, + 0.004550992976874113, + 0.004121947567909956, + 0.03247039392590523, + 0.05479702353477478, + 0.025020848959684372, + 0.028433972969651222, + 0.0314236655831337, + 0.04004845768213272, + 0.02113998867571354, + 0.025043878704309464, + 0.04423002153635025, + 0.011964299716055393, + 0.008182695135474205, + 0.005102252587676048, + 0.011540929786860943, + 0.05539459362626076, + 0.006330587435513735, + 0.008480021730065346, + 0.01823713257908821, + 0.007548865396529436, + 0.016212014481425285, + 0.008389156311750412, + 0.008639953099191189, + 0.01882251910865307, + 0.005616846028715372, + 0.020050229504704475, + 0.007094454485923052, + 0.005251157563179731, + 0.006949970964342356, + 0.005076237954199314, + 0.004671595990657806, + 0.004487006459385157, + 0.002993868663907051, + 0.0048570577055215836, + 0.004306148737668991, + 0.0051171560771763325, + 0.007342514116317034, + 0.005416735541075468, + 0.0032361198682338, + 0.0028993526939302683, + 0.002360384212806821, + 0.003055193927139044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.01421845331788063 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739622913300991 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739622913300991, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739622913300991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09365800768136978, + 0.19398494064807892, + 0.04027194529771805, + 0.04750915616750717, + 0.12167146801948547, + 0.03570074215531349, + 0.32308515906333923, + 0.0800778791308403, + 0.027338538318872452, + 0.03670218214392662 + ], + "credit_map": [ + 0.01299253385514021, + 0.00974602997303009, + 0.0063322377391159534, + 0.009586683474481106, + 0.009158249944448471, + 0.016155622899532318, + 0.009385352022945881, + 0.020301295444369316, + 0.011034693568944931, + 0.0074720545671880245, + 0.009979929774999619, + 0.007087651174515486, + 0.008349071256816387, + 0.06716587394475937, + 0.014041287824511528, + 0.06885436177253723, + 0.005870623514056206, + 0.00463970797136426, + 0.0047238669358193874, + 0.0026544590946286917, + 0.005986736621707678, + 0.005608897190541029, + 0.005427730269730091, + 0.005359923467040062, + 0.01361132599413395, + 0.004253333900123835, + 0.00553315132856369, + 0.004562740679830313, + 0.003166760317981243, + 0.006974367890506983, + 0.003438755637034774, + 0.005968715529888868, + 0.019913718104362488, + 0.015782028436660767, + 0.017889084294438362, + 0.017883211374282837, + 0.010623299516737461, + 0.01739826425909996, + 0.008256141096353531, + 0.013925710693001747, + 0.0043563563376665115, + 0.0071791489608585835, + 0.00422970624640584, + 0.0032427774276584387, + 0.004981844685971737, + 0.005210869945585728, + 0.004017779137939215, + 0.002482257317751646, + 0.11368606239557266, + 0.01120160985738039, + 0.013831689022481441, + 0.005305401980876923, + 0.02273339405655861, + 0.13672029972076416, + 0.005137799307703972, + 0.014468889683485031, + 0.014543650671839714, + 0.006617977283895016, + 0.006959243677556515, + 0.004305895883589983, + 0.006522119976580143, + 0.017023233696818352, + 0.004838528577238321, + 0.019267216324806213, + 0.004019075073301792, + 0.0033496811520308256, + 0.004812030587345362, + 0.0029144156724214554, + 0.0019547624979168177, + 0.002965780673548579, + 0.002178001683205366, + 0.005144789814949036, + 0.004789875354617834, + 0.0051390500739216805, + 0.008280551061034203, + 0.00431442353874445, + 0.004125264473259449, + 0.003395036794245243, + 0.002479010261595249, + 0.004178966861218214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.006739622913300991 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370674327015877 + ], + "opposite_logprob_tokens": [ + -3.4087066650390625 + ], + "logprob_reward": -0.03370674327015877, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370674327015877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09137339890003204, + 0.25270935893058777, + 0.06986089795827866, + 0.05768461525440216, + 0.05483947694301605, + 0.04176558181643486, + 0.10351864248514175, + 0.13190896809101105, + 0.1425689160823822, + 0.05377015843987465 + ], + "credit_map": [ + 0.013088815845549107, + 0.0074250539764761925, + 0.005175563506782055, + 0.005809841211885214, + 0.0048667811788618565, + 0.023663554340600967, + 0.006168396212160587, + 0.02517540194094181, + 0.01608552783727646, + 0.011696187779307365, + 0.016165167093276978, + 0.012220625765621662, + 0.010640040971338749, + 0.08131366968154907, + 0.022841155529022217, + 0.08174697309732437, + 0.01506822183728218, + 0.008116638287901878, + 0.007425150368362665, + 0.004683620296418667, + 0.009317819960415363, + 0.01004761178046465, + 0.006070674397051334, + 0.009131167083978653, + 0.008910991251468658, + 0.005678300280123949, + 0.008008521981537342, + 0.0052279578521847725, + 0.006480232812464237, + 0.008237733505666256, + 0.008751020766794682, + 0.006389854941517115, + 0.007223899010568857, + 0.00798310711979866, + 0.005370844155550003, + 0.006480432581156492, + 0.007537694647908211, + 0.00903107039630413, + 0.005488211754709482, + 0.0057242196053266525, + 0.0052562132477760315, + 0.006728084292262793, + 0.00511419540271163, + 0.004817705135792494, + 0.005827622022479773, + 0.0058782994747161865, + 0.004250642843544483, + 0.003892816137522459, + 0.026937460526823997, + 0.011225158348679543, + 0.006356248632073402, + 0.00437102559953928, + 0.008839504793286324, + 0.03426368907094002, + 0.005977966357022524, + 0.005547583103179932, + 0.020252887159585953, + 0.013336197473108768, + 0.012808435596525669, + 0.010165158659219742, + 0.012633858248591423, + 0.02278417907655239, + 0.009294239804148674, + 0.030634012073278427, + 0.022227739915251732, + 0.022479822859168053, + 0.02550847828388214, + 0.020071720704436302, + 0.013679477386176586, + 0.013552781194448471, + 0.010497082024812698, + 0.014551802538335323, + 0.008282458409667015, + 0.0063956682570278645, + 0.00873310025781393, + 0.004996451549232006, + 0.005872679874300957, + 0.0047943685203790665, + 0.004984543193131685, + 0.009710890240967274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.03370674327015877 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level close-up of a massive great white shark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shark's body colored with a gradient of dark slate gray on top and white on the underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment an endless expanse of deep cobalt blue water with faint sunlight penetrating from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the shark's left eye a black, unblinking orb fixed directly on the lens?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera exhibit a subtle handheld tremor while the shark remains largely motionless?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705424427986145 + ], + "opposite_logprob_tokens": [ + -2.127054214477539 + ], + "logprob_reward": -0.12705424427986145, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705424427986145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11969351768493652, + 0.1050797551870346, + 0.03554221987724304, + 0.04052649438381195, + 0.04613661393523216, + 0.03309476375579834, + 0.310265451669693, + 0.14287883043289185, + 0.13854293525218964, + 0.02823939360678196 + ], + "credit_map": [ + 0.020320065319538116, + 0.009780912660062313, + 0.00746723311021924, + 0.006224737502634525, + 0.007199966814368963, + 0.028510257601737976, + 0.010792526416480541, + 0.029397811740636826, + 0.00821775384247303, + 0.006209668703377247, + 0.007107353303581476, + 0.005405576433986425, + 0.0042950366623699665, + 0.03345704451203346, + 0.006523371674120426, + 0.03386394679546356, + 0.00751973083242774, + 0.0034581399522721767, + 0.004562679212540388, + 0.0019419632153585553, + 0.0034559760242700577, + 0.00565820187330246, + 0.002640484133735299, + 0.006305046379566193, + 0.005838913377374411, + 0.0034180080983787775, + 0.005121349822729826, + 0.0052261147648096085, + 0.004469050094485283, + 0.005476268008351326, + 0.005750331562012434, + 0.005226457957178354, + 0.005665455013513565, + 0.005618465598672628, + 0.005616573616862297, + 0.005778043530881405, + 0.005568194203078747, + 0.007036216091364622, + 0.0047404407523572445, + 0.006113228853791952, + 0.004835046362131834, + 0.006172493565827608, + 0.0035956709180027246, + 0.002872068202123046, + 0.004235593602061272, + 0.0043038781732320786, + 0.002751791151240468, + 0.004328221548348665, + 0.07842348515987396, + 0.033350326120853424, + 0.019215509295463562, + 0.013688934966921806, + 0.02845410630106926, + 0.09770717471837997, + 0.021046366542577744, + 0.01837954856455326, + 0.023869382217526436, + 0.01752443239092827, + 0.023412004113197327, + 0.01760692521929741, + 0.011768761090934277, + 0.02268068864941597, + 0.008287075906991959, + 0.017729565501213074, + 0.021596431732177734, + 0.01303776539862156, + 0.024431584402918816, + 0.012577938847243786, + 0.010988702066242695, + 0.017100950703024864, + 0.010868694633245468, + 0.02794087864458561, + 0.004194512497633696, + 0.0038669679779559374, + 0.0037175596226006746, + 0.0036726996768265963, + 0.003083917312324047, + 0.0029758960008621216, + 0.0023333497811108828, + 0.00439448980614543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.12705424427986145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_5/index_tensor([5615, 5615])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_5/index_tensor([5615, 5615])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..aa08b7a663fc1b6a527f2220bb631431290a188a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_5/index_tensor([5615, 5615])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An initially static, extreme wide shot slowly tilts right, tracking a mesmerizing natural event across a vast meadow bathed in the warm, golden afternoon sun. A dense, shimmering cloud of hawthorn seeds drifts effortlessly on the breeze. Each seed, a tiny black achene, hovers atop a delicate, fluffy white nipple, catching the low-angle sunlight, causing the entire swarm to emit an ethereal, backlit glow. They soar, turn, and tumble in a slow, chaotic, yet graceful dance, some rising high against the soft blue sky while others skim the tips of tall green grass and scattered wildflowers below. The vast field stretches to the hazy, distant treeline, where long, soft shadows cast by the setting sun accentuate the texture of the landscape. The overall atmosphere is one of profound tranquility and dreamlike wonder, capturing a fleeting, magical moment in nature.", + "seed": 19827, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_5/index_5615_seed_19827_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18431165330111982, + "vlm_logprob": -0.18431165330111982, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.18431165330111982, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4789806604385376 + ], + "opposite_logprob_tokens": [ + -0.9789806604385376 + ], + "logprob_reward": -0.4789806604385376, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4789806604385376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12898051738739014, + 0.2718338966369629, + 0.03600986674427986, + 0.18851114809513092, + 0.045249197632074356, + 0.09605646878480911, + 0.056193578988313675, + 0.017992934212088585, + 0.08229115605354309, + 0.0768812894821167 + ], + "credit_map": [ + 0.006906671915203333, + 0.011163268238306046, + 0.01498539187014103, + 0.01862817443907261, + 0.012410078197717667, + 0.01273091696202755, + 0.028754068538546562, + 0.023401940241456032, + 0.00568500068038702, + 0.007170049473643303, + 0.01847868412733078, + 0.02252875454723835, + 0.009686506353318691, + 0.007545188069343567, + 0.11565400660037994, + 0.08508572727441788, + 0.0020865893457084894, + 0.006761085242033005, + 0.004025632049888372, + 0.005438980180770159, + 0.004435473121702671, + 0.0035962755791842937, + 0.003794722259044647, + 0.00587110873311758, + 0.008024816401302814, + 0.006626952439546585, + 0.015321196988224983, + 0.014664873480796814, + 0.008507405407726765, + 0.012754874303936958, + 0.06831568479537964, + 0.05429533123970032, + 0.0038641667924821377, + 0.003820322686806321, + 0.007956480607390404, + 0.006047883536666632, + 0.0037343124859035015, + 0.004631160292774439, + 0.007454993668943644, + 0.007739876862615347, + 0.0071431915275752544, + 0.00835160817950964, + 0.01622663624584675, + 0.012879722751677036, + 0.006056951824575663, + 0.010667686350643635, + 0.011369062587618828, + 0.02336161583662033, + 0.004218643065541983, + 0.00287839537486434, + 0.010536129586398602, + 0.010477591305971146, + 0.0059938980266451836, + 0.006081992294639349, + 0.006631226744502783, + 0.009375703521072865, + 0.001230863039381802, + 0.0017410488799214363, + 0.0021573486737906933, + 0.0015893197851255536, + 0.0029117141384631395, + 0.003583562094718218, + 0.002382840495556593, + 0.0023962375707924366, + 0.0071597956120967865, + 0.006671796552836895, + 0.014514215290546417, + 0.010033367201685905, + 0.00930240098387003, + 0.01177438348531723, + 0.011387837119400501, + 0.011447363533079624, + 0.004036734811961651, + 0.005773230921477079, + 0.020330028608441353, + 0.009559009224176407, + 0.006075129844248295, + 0.0059084203094244, + 0.007907522842288017, + 0.01729121431708336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.4789806604385376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330239862203598 + ], + "opposite_logprob_tokens": [ + -3.7733023166656494 + ], + "logprob_reward": -0.02330239862203598, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330239862203598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22180330753326416, + 0.26946914196014404, + 0.042326975613832474, + 0.1760169267654419, + 0.02492167428135872, + 0.03402754291892052, + 0.021920230239629745, + 0.02422604151070118, + 0.07542208582162857, + 0.10986609756946564 + ], + "credit_map": [ + 0.011305262334644794, + 0.019678182899951935, + 0.02169010415673256, + 0.02517678029835224, + 0.019751664251089096, + 0.022318607196211815, + 0.060666002333164215, + 0.041216712445020676, + 0.006884842645376921, + 0.008954932913184166, + 0.020205514505505562, + 0.022774064913392067, + 0.008879154920578003, + 0.0090709263458848, + 0.11509376764297485, + 0.07760593295097351, + 0.003337826579809189, + 0.005497208796441555, + 0.004521716386079788, + 0.0066495356149971485, + 0.006799975410103798, + 0.004957939498126507, + 0.00618613651022315, + 0.004376636818051338, + 0.007253714371472597, + 0.00431024981662631, + 0.01990712806582451, + 0.01467977836728096, + 0.006427302025258541, + 0.007373100612312555, + 0.06791799515485764, + 0.04814765229821205, + 0.0017568847397342324, + 0.0025153697934001684, + 0.0036695334129035473, + 0.004584347363561392, + 0.0018403120338916779, + 0.002794107422232628, + 0.004547923803329468, + 0.00321319536305964, + 0.002819585846737027, + 0.0031292112544178963, + 0.00422504311427474, + 0.006160725373774767, + 0.002603944158181548, + 0.00423610582947731, + 0.005030215717852116, + 0.005822709761559963, + 0.001133244950324297, + 0.0011791366850957274, + 0.004026161041110754, + 0.003326605772599578, + 0.0036800766829401255, + 0.0033730696886777878, + 0.0021908357739448547, + 0.0030111009255051613, + 0.0015389967011287808, + 0.0028048348613083363, + 0.0035592326894402504, + 0.004240191541612148, + 0.0017365575768053532, + 0.0030163927003741264, + 0.002987043932080269, + 0.00434279115870595, + 0.0058522578328847885, + 0.0046652499586343765, + 0.013322710059583187, + 0.009214892983436584, + 0.007673068903386593, + 0.011561732739210129, + 0.012273155152797699, + 0.010859024710953236, + 0.007390522863715887, + 0.009519224986433983, + 0.01630895957350731, + 0.014013497158885002, + 0.012886174954473972, + 0.01744699850678444, + 0.0163466464728117, + 0.015954071655869484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02330239862203598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871375322341919 + ], + "opposite_logprob_tokens": [ + -1.137137532234192 + ], + "logprob_reward": -0.3871375322341919, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871375322341919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2892986238002777, + 0.2680375576019287, + 0.07881083339452744, + 0.06605400145053864, + 0.024320872500538826, + 0.04920198395848274, + 0.02248104102909565, + 0.023239189758896828, + 0.09982339292764664, + 0.07873249053955078 + ], + "credit_map": [ + 0.013163294643163681, + 0.02413015626370907, + 0.027779895812273026, + 0.03243444487452507, + 0.023144926875829697, + 0.023869983851909637, + 0.08799141645431519, + 0.056784488260746, + 0.005638842470943928, + 0.008169637992978096, + 0.016763288527727127, + 0.020491329953074455, + 0.010233054868876934, + 0.006308934651315212, + 0.1127161756157875, + 0.08771631121635437, + 0.0036662681959569454, + 0.007919195108115673, + 0.007123282179236412, + 0.010099448263645172, + 0.010738959535956383, + 0.005905898753553629, + 0.022737303748726845, + 0.010620472021400928, + 0.004878019914031029, + 0.006103275343775749, + 0.008252532221376896, + 0.011796030215919018, + 0.0054591018706560135, + 0.01142432913184166, + 0.007520992774516344, + 0.010619711130857468, + 0.0021651596762239933, + 0.002051812130957842, + 0.003715568222105503, + 0.004202097188681364, + 0.0020827664993703365, + 0.002704298123717308, + 0.004183848388493061, + 0.003215319477021694, + 0.0024931798689067364, + 0.002568757627159357, + 0.011739996261894703, + 0.007625126745551825, + 0.0024473306257277727, + 0.004914138000458479, + 0.005913033615797758, + 0.01150041539222002, + 0.0014692234108224511, + 0.0013816399732604623, + 0.002681852551177144, + 0.002270002383738756, + 0.003010484855622053, + 0.004412461072206497, + 0.0027675628662109375, + 0.004487812053412199, + 0.0019381589954718947, + 0.00321363378316164, + 0.0030230050906538963, + 0.004135665483772755, + 0.0019477704772725701, + 0.003425170900300145, + 0.0025272623170167208, + 0.003028522478416562, + 0.007543494459241629, + 0.0068307602778077126, + 0.01766446977853775, + 0.011468698270618916, + 0.009702309966087341, + 0.014901767484843731, + 0.015956304967403412, + 0.015755580738186836, + 0.005377677734941244, + 0.005437180399894714, + 0.016599252820014954, + 0.010640296153724194, + 0.006547159980982542, + 0.010576769709587097, + 0.011330552399158478, + 0.012223600409924984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.3871375322341919 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077635809779167 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.016077635809779167, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077635809779167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2557907998561859, + 0.2941989302635193, + 0.08930507302284241, + 0.034295205026865005, + 0.016628321260213852, + 0.058200035244226456, + 0.029455015435814857, + 0.02590613253414631, + 0.12069440633058548, + 0.07552603632211685 + ], + "credit_map": [ + 0.011067957617342472, + 0.018635757267475128, + 0.026998713612556458, + 0.032377034425735474, + 0.0200918298214674, + 0.021842647343873978, + 0.0734124481678009, + 0.05136439949274063, + 0.006175802554935217, + 0.006789009552448988, + 0.020426485687494278, + 0.023434853181242943, + 0.009170814417302608, + 0.007828484289348125, + 0.1268688440322876, + 0.09350463002920151, + 0.004147733096033335, + 0.007791679352521896, + 0.007969886064529419, + 0.010457429103553295, + 0.00991335604339838, + 0.006028290838003159, + 0.029287122189998627, + 0.013709579594433308, + 0.00267758104018867, + 0.002987559651955962, + 0.004183000419288874, + 0.005092672538012266, + 0.002243622438982129, + 0.003949252888560295, + 0.0074297040700912476, + 0.005731811746954918, + 0.0011520151747390628, + 0.0015763003611937165, + 0.0032870080322027206, + 0.0026764729991555214, + 0.0012910286895930767, + 0.0017206795746460557, + 0.0021369336172938347, + 0.0027878833934664726, + 0.002534274710342288, + 0.0031034459825605154, + 0.016144124791026115, + 0.009248151443898678, + 0.002816463354974985, + 0.004929596092551947, + 0.005180462263524532, + 0.014243519864976406, + 0.0016357391141355038, + 0.0013348760548979044, + 0.0042716446332633495, + 0.003379786852747202, + 0.003441751701757312, + 0.005328011699020863, + 0.0035708355717360973, + 0.006492369808256626, + 0.002093392191454768, + 0.0033824003767222166, + 0.0034690594766288996, + 0.004318192135542631, + 0.002176907379180193, + 0.004105777014046907, + 0.002985588274896145, + 0.0033748161513358355, + 0.009559810161590576, + 0.007733990903943777, + 0.02200300060212612, + 0.014595413580536842, + 0.011823592707514763, + 0.018771503120660782, + 0.018523143604397774, + 0.017683953046798706, + 0.00453641451895237, + 0.004639675375074148, + 0.020600805059075356, + 0.008585667237639427, + 0.0051419781520962715, + 0.007652003783732653, + 0.00823392253369093, + 0.016135573387145996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.016077635809779167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016060039401054382 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.016060039401054382, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016060039401054382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12423624843358994, + 0.3610498607158661, + 0.05248066037893295, + 0.2479935586452484, + 0.02701534889638424, + 0.024020133540034294, + 0.011189625598490238, + 0.056618157774209976, + 0.015445391647517681, + 0.07995101809501648 + ], + "credit_map": [ + 0.0059529803693294525, + 0.010277451947331429, + 0.011250191368162632, + 0.013226966373622417, + 0.008471676148474216, + 0.01140930037945509, + 0.03841588646173477, + 0.025231793522834778, + 0.008637155406177044, + 0.019821498543024063, + 0.02195645123720169, + 0.028132779523730278, + 0.019636211916804314, + 0.01422650832682848, + 0.1402074545621872, + 0.10843179374933243, + 0.002416698494926095, + 0.003795288037508726, + 0.005063067190349102, + 0.005323350429534912, + 0.002626304281875491, + 0.002360166050493717, + 0.019461212679743767, + 0.01143457368016243, + 0.01329644862562418, + 0.01098380796611309, + 0.019550686702132225, + 0.016702108085155487, + 0.014104286208748817, + 0.027827877551317215, + 0.07490222156047821, + 0.07062613219022751, + 0.0028912341222167015, + 0.0029440773651003838, + 0.003797453362494707, + 0.004785478115081787, + 0.002047359710559249, + 0.003572253044694662, + 0.003322218544781208, + 0.0036552739329636097, + 0.001653414685279131, + 0.0022779402788728476, + 0.005735199432820082, + 0.0032725322525948286, + 0.0014423785032704473, + 0.0017054309137165546, + 0.002338849939405918, + 0.005594387650489807, + 0.0007894322625361383, + 0.0006797611713409424, + 0.0017106598243117332, + 0.001284051570110023, + 0.0013215006329119205, + 0.0022512050345540047, + 0.0013183081755414605, + 0.0018347057048231363, + 0.0033943962771445513, + 0.00616559199988842, + 0.006584738846868277, + 0.005209012422710657, + 0.007845192216336727, + 0.012719136662781239, + 0.008446970954537392, + 0.0062531232833862305, + 0.0015988540835678577, + 0.001468714326620102, + 0.003347850637510419, + 0.0023780260235071182, + 0.0015627251705154777, + 0.001411592704243958, + 0.0018986475188285112, + 0.0017789803678169847, + 0.007775983773171902, + 0.012433460913598537, + 0.013008926063776016, + 0.00927809625864029, + 0.009127668105065823, + 0.007317004259675741, + 0.012083163484930992, + 0.008926707319915295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.016060039401054382 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An initially static, extreme wide shot slowly tilts right, tracking a mesmerizing natural event across a vast meadow bathed in the warm, golden afternoon sun. A dense, shimmering cloud of hawthorn seeds drifts effortlessly on the breeze. Each seed, a tiny black achene, hovers atop a delicate, fluffy white nipple, catching the low-angle sunlight, causing the entire swarm to emit an ethereal, backlit glow. They soar, turn, and tumble in a slow, chaotic, yet graceful dance, some rising high against the soft blue sky while others skim the tips of tall green grass and scattered wildflowers below. The vast field stretches to the hazy, distant treeline, where long, soft shadows cast by the setting sun accentuate the texture of the landscape. The overall atmosphere is one of profound tranquility and dreamlike wonder, capturing a fleeting, magical moment in nature.", + "seed": 19828, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_5/index_5615_seed_19828_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09474991140887142, + "vlm_logprob": -0.09474991140887142, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.09474991140887142, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21817517280578613 + ], + "opposite_logprob_tokens": [ + -1.7181751728057861 + ], + "logprob_reward": -0.21817517280578613, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21817517280578613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04524565860629082, + 0.2785591185092926, + 0.04313146322965622, + 0.1550818830728531, + 0.0548674538731575, + 0.248104065656662, + 0.03390306234359741, + 0.09495171904563904, + 0.024218816310167313, + 0.021936699748039246 + ], + "credit_map": [ + 0.004239845089614391, + 0.006350474897772074, + 0.005192122887820005, + 0.007174900267273188, + 0.006433832459151745, + 0.003302633995190263, + 0.007794253062456846, + 0.004757601302117109, + 0.010192405432462692, + 0.010412724688649178, + 0.012580886483192444, + 0.014378848485648632, + 0.014059516601264477, + 0.00892251543700695, + 0.11002986133098602, + 0.09798236191272736, + 0.005373802967369556, + 0.0032957468647509813, + 0.00645854277536273, + 0.004881810396909714, + 0.0031704402063041925, + 0.0027880799025297165, + 0.011478010565042496, + 0.0056850360706448555, + 0.008379324339330196, + 0.006894814781844616, + 0.027198903262615204, + 0.03531767427921295, + 0.016110673546791077, + 0.017909567803144455, + 0.022512679919600487, + 0.02075825445353985, + 0.005378139205276966, + 0.005394240841269493, + 0.011435543186962605, + 0.007270914036780596, + 0.00659500015899539, + 0.005403363145887852, + 0.007095458917319775, + 0.006294799502938986, + 0.020053531974554062, + 0.024250607937574387, + 0.052004292607307434, + 0.04488203674554825, + 0.019979393109679222, + 0.025776322931051254, + 0.030132878571748734, + 0.031025024130940437, + 0.003450034884735942, + 0.002285231137648225, + 0.003834417089819908, + 0.00512399896979332, + 0.0050487513653934, + 0.0035864871460944414, + 0.004158528055995703, + 0.006415617652237415, + 0.008950166404247284, + 0.007128940429538488, + 0.012028750032186508, + 0.006072866730391979, + 0.012342301197350025, + 0.026780130341649055, + 0.009309343062341213, + 0.012339233420789242, + 0.001889036619104445, + 0.002033041324466467, + 0.003096214961260557, + 0.004656991921365261, + 0.0023896254133433104, + 0.0016317981062456965, + 0.0050371079705655575, + 0.0034850025549530983, + 0.0016720988787710667, + 0.0021289181895554066, + 0.003348034108057618, + 0.003803541650995612, + 0.0020753981079906225, + 0.0020326205994933844, + 0.003671037033200264, + 0.0032050537411123514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.21817517280578613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009801830165088177 + ], + "opposite_logprob_tokens": [ + -4.634801864624023 + ], + "logprob_reward": -0.009801830165088177, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009801830165088177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07382771372795105, + 0.6865739226341248, + 0.03407149761915207, + 0.0831836462020874, + 0.014959732070565224, + 0.036773182451725006, + 0.010426688008010387, + 0.018474826589226723, + 0.026752758771181107, + 0.01495592575520277 + ], + "credit_map": [ + 0.006424376741051674, + 0.010584672912955284, + 0.011386225000023842, + 0.011743639595806599, + 0.008464181795716286, + 0.0058992775157094, + 0.009903308004140854, + 0.009422041475772858, + 0.027507003396749496, + 0.03129660710692406, + 0.031191039830446243, + 0.035209279507398605, + 0.03641205653548241, + 0.02443569339811802, + 0.25837403535842896, + 0.2421482652425766, + 0.003826497122645378, + 0.0018356818472966552, + 0.0046866899356245995, + 0.004408259876072407, + 0.002954778028652072, + 0.0021521945018321276, + 0.008904737420380116, + 0.005302661098539829, + 0.0059359208680689335, + 0.003772416617721319, + 0.012367875315248966, + 0.023601625114679337, + 0.003932577092200518, + 0.0034111356362700462, + 0.017030207440257072, + 0.013131891377270222, + 0.0014480117242783308, + 0.001554662361741066, + 0.0016170424642041326, + 0.002089061075821519, + 0.003046521684154868, + 0.0014936403604224324, + 0.002015473088249564, + 0.0016953201266005635, + 0.002641899511218071, + 0.003547756001353264, + 0.00788422953337431, + 0.006391580682247877, + 0.00296634528785944, + 0.0039795925840735435, + 0.004628052935004234, + 0.004733728710561991, + 0.0011865959968417883, + 0.0011118724942207336, + 0.001363289193250239, + 0.0010519042843952775, + 0.001227897242642939, + 0.001313547370955348, + 0.0015265835681930184, + 0.0016449984395876527, + 0.0021458561532199383, + 0.0014989505289122462, + 0.002490963088348508, + 0.002297463361173868, + 0.0016984875546768308, + 0.0024210179690271616, + 0.002524190815165639, + 0.003397897584363818, + 0.001970030600205064, + 0.001963029382750392, + 0.0035366187803447247, + 0.0046198200434446335, + 0.0029455290641635656, + 0.0024103692267090082, + 0.005571315065026283, + 0.0037360484711825848, + 0.0011915029026567936, + 0.0008754120790399611, + 0.0018611024133861065, + 0.002046401146799326, + 0.0022326535545289516, + 0.0021186277735978365, + 0.0024419466499239206, + 0.0021882792934775352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.009801830165088177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561071813106537 + ], + "opposite_logprob_tokens": [ + -1.6006107330322266 + ], + "logprob_reward": -0.22561071813106537, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561071813106537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13137812912464142, + 0.436351478099823, + 0.06156545877456665, + 0.06924386322498322, + 0.03633420169353485, + 0.046406760811805725, + 0.041164327412843704, + 0.0644274652004242, + 0.08267287164926529, + 0.03045547939836979 + ], + "credit_map": [ + 0.010791880078613758, + 0.014458924531936646, + 0.020326871424913406, + 0.01858462207019329, + 0.017911948263645172, + 0.007392325438559055, + 0.024246715009212494, + 0.017664827406406403, + 0.01573825813829899, + 0.013210108503699303, + 0.019322002306580544, + 0.02387230098247528, + 0.019541919231414795, + 0.01392060425132513, + 0.1748172789812088, + 0.15592898428440094, + 0.005894533358514309, + 0.005170911084860563, + 0.009665087796747684, + 0.009691665880382061, + 0.006939082406461239, + 0.0064577264711260796, + 0.007898843847215176, + 0.009847607463598251, + 0.006159014999866486, + 0.0044672368094325066, + 0.008885979652404785, + 0.015410141088068485, + 0.00531820859760046, + 0.005256462376564741, + 0.012553203850984573, + 0.011193612590432167, + 0.0027323735412210226, + 0.0026901422534137964, + 0.005323972087353468, + 0.007120657246559858, + 0.0025365916080772877, + 0.004281069152057171, + 0.00698834890499711, + 0.0046610478311777115, + 0.0038808139506727457, + 0.005838321056216955, + 0.006390845403075218, + 0.005740945227444172, + 0.004554836079478264, + 0.004518845118582249, + 0.009038693271577358, + 0.006443462334573269, + 0.003428148804232478, + 0.0025018672458827496, + 0.005349913612008095, + 0.006158576346933842, + 0.006384450010955334, + 0.003914063796401024, + 0.0073613268323242664, + 0.0060659791342914104, + 0.0054947445169091225, + 0.005279775243252516, + 0.008094919845461845, + 0.007079209666699171, + 0.007247321307659149, + 0.010448532178997993, + 0.00878980290144682, + 0.011993160471320152, + 0.006843435112386942, + 0.008456307463347912, + 0.010252364911139011, + 0.013402444310486317, + 0.0096110999584198, + 0.0061206212267279625, + 0.01716420240700245, + 0.010822394862771034, + 0.002458681119605899, + 0.0039405482821166515, + 0.0036785711999982595, + 0.005630480125546455, + 0.0038172886706888676, + 0.0020965358708053827, + 0.005312989000231028, + 0.0035203825682401657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.22561071813106537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630014326423407 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007630014326423407, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630014326423407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14449603855609894, + 0.46982112526893616, + 0.06422638148069382, + 0.05667719617486, + 0.03613132983446121, + 0.0334295853972435, + 0.03131013363599777, + 0.053516704589128494, + 0.0733385980129242, + 0.037052955478429794 + ], + "credit_map": [ + 0.010927223600447178, + 0.015062891878187656, + 0.021394191309809685, + 0.022885974496603012, + 0.018155407160520554, + 0.008453158661723137, + 0.02935139834880829, + 0.01826578937470913, + 0.01781642995774746, + 0.013981133699417114, + 0.019935160875320435, + 0.025841448456048965, + 0.019802052527666092, + 0.015222889371216297, + 0.18573272228240967, + 0.17148926854133606, + 0.007085913792252541, + 0.005012950394302607, + 0.010129110887646675, + 0.008846605196595192, + 0.0068210819736123085, + 0.007456574123352766, + 0.008282867260277271, + 0.010591276921331882, + 0.006915183272212744, + 0.004674182739108801, + 0.00685096113011241, + 0.009087887592613697, + 0.005842356011271477, + 0.005888637155294418, + 0.009143651463091373, + 0.008274329826235771, + 0.002513582119718194, + 0.0024562329053878784, + 0.005684452131390572, + 0.007485412526875734, + 0.0025984805542975664, + 0.003689088160172105, + 0.006659654900431633, + 0.0050444211810827255, + 0.002530609490349889, + 0.003529973793774843, + 0.004263587296009064, + 0.004656143952161074, + 0.004068617708981037, + 0.0048081399872899055, + 0.005335384514182806, + 0.0042371260933578014, + 0.0021991885732859373, + 0.00431870948523283, + 0.004172402434051037, + 0.003915904555469751, + 0.0036244813818484545, + 0.0028145096730440855, + 0.004493534564971924, + 0.005771398078650236, + 0.005304467864334583, + 0.004045312292873859, + 0.007126842625439167, + 0.006598995998501778, + 0.005200183484703302, + 0.008376664482057095, + 0.008011538535356522, + 0.008852694183588028, + 0.006045487709343433, + 0.00800046231597662, + 0.009667661972343922, + 0.011435977183282375, + 0.008954928256571293, + 0.005339760333299637, + 0.01453589741140604, + 0.009358419105410576, + 0.003553351853042841, + 0.004370285663753748, + 0.00468513322994113, + 0.007467001210898161, + 0.00376018020324409, + 0.0024434200022369623, + 0.006048929877579212, + 0.004724650643765926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.007630014326423407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012531821615993977 + ], + "opposite_logprob_tokens": [ + -4.387531757354736 + ], + "logprob_reward": -0.012531821615993977, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012531821615993977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10451557487249374, + 0.4241138994693756, + 0.07279260456562042, + 0.11232461035251617, + 0.04021801799535751, + 0.06412561982870102, + 0.018334953114390373, + 0.031066110357642174, + 0.06582848727703094, + 0.06668005883693695 + ], + "credit_map": [ + 0.008206543512642384, + 0.012783619575202465, + 0.016598548740148544, + 0.017771290615200996, + 0.011334930546581745, + 0.007638679817318916, + 0.017710071057081223, + 0.012471904046833515, + 0.016063950955867767, + 0.012348989956080914, + 0.018048709258437157, + 0.020160172134637833, + 0.01506798155605793, + 0.013456192798912525, + 0.16962753236293793, + 0.15934039652347565, + 0.00602804496884346, + 0.006675987038761377, + 0.007476126309484243, + 0.007323915138840675, + 0.005953378509730101, + 0.004458625800907612, + 0.023273032158613205, + 0.011603505350649357, + 0.004848152864724398, + 0.004046743270009756, + 0.017813311889767647, + 0.028575627133250237, + 0.009640192613005638, + 0.011204284615814686, + 0.019531279802322388, + 0.0166650228202343, + 0.004528920631855726, + 0.004166930913925171, + 0.006437278352677822, + 0.003813011571764946, + 0.008258506655693054, + 0.004999623168259859, + 0.004483636002987623, + 0.0035301141906529665, + 0.00551975704729557, + 0.005573684815317392, + 0.01357568521052599, + 0.009292533621191978, + 0.0057788207195699215, + 0.008661267347633839, + 0.008502460084855556, + 0.007221415638923645, + 0.0017021136591210961, + 0.0012431828072294593, + 0.0023325984366238117, + 0.003192547243088484, + 0.0020257006399333477, + 0.0014034254709258676, + 0.0027548614889383316, + 0.003680524881929159, + 0.0042664348147809505, + 0.0023248642683029175, + 0.004960666876286268, + 0.0030549154616892338, + 0.004166735336184502, + 0.0045228819362819195, + 0.0034729191102087498, + 0.004296693950891495, + 0.005044244695454836, + 0.007817975245416164, + 0.009312501177191734, + 0.008234268985688686, + 0.00616793055087328, + 0.006597612984478474, + 0.014000925235450268, + 0.008653038181364536, + 0.009725499898195267, + 0.010677659884095192, + 0.007637349888682365, + 0.00779132591560483, + 0.008882858790457249, + 0.0052975984290242195, + 0.008333100937306881, + 0.008334667421877384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.012531821615993977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An initially static, extreme wide shot slowly tilts right, tracking a mesmerizing natural event across a vast meadow bathed in the warm, golden afternoon sun. A dense, shimmering cloud of hawthorn seeds drifts effortlessly on the breeze. Each seed, a tiny black achene, hovers atop a delicate, fluffy white nipple, catching the low-angle sunlight, causing the entire swarm to emit an ethereal, backlit glow. They soar, turn, and tumble in a slow, chaotic, yet graceful dance, some rising high against the soft blue sky while others skim the tips of tall green grass and scattered wildflowers below. The vast field stretches to the hazy, distant treeline, where long, soft shadows cast by the setting sun accentuate the texture of the landscape. The overall atmosphere is one of profound tranquility and dreamlike wonder, capturing a fleeting, magical moment in nature.", + "seed": 19829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_5/index_5615_seed_19829_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1182269879616797, + "vlm_logprob": -0.1182269879616797, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.1182269879616797, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17548047006130219 + ], + "opposite_logprob_tokens": [ + -1.9254804849624634 + ], + "logprob_reward": -0.17548047006130219, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17548047006130219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09597693383693695, + 0.029604140669107437, + 0.06199837103486061, + 0.04825909063220024, + 0.06624472141265869, + 0.07015905529260635, + 0.5186715722084045, + 0.03672107309103012, + 0.033067237585783005, + 0.03929780423641205 + ], + "credit_map": [ + 0.007519324775785208, + 0.009473017416894436, + 0.019666267558932304, + 0.019253762438893318, + 0.01334530021995306, + 0.008233631029725075, + 0.010441242717206478, + 0.008044383488595486, + 0.002852465957403183, + 0.0039027200546115637, + 0.003387727774679661, + 0.004709101282060146, + 0.0055979518219828606, + 0.002865226473659277, + 0.0032069480512291193, + 0.00308199692517519, + 0.00496356375515461, + 0.00896353181451559, + 0.004438973497599363, + 0.007186924573034048, + 0.018995502963662148, + 0.005754630547016859, + 0.006543232128024101, + 0.005152004770934582, + 0.004167681094259024, + 0.004967879503965378, + 0.0042974017560482025, + 0.009009867906570435, + 0.009857318364083767, + 0.005681620445102453, + 0.00631463248282671, + 0.0039626904763281345, + 0.004774041008204222, + 0.005056680645793676, + 0.010411602444946766, + 0.010889207944273949, + 0.00615114439278841, + 0.006169903092086315, + 0.014007785357534885, + 0.008784346282482147, + 0.008176570758223534, + 0.013801183551549911, + 0.012425294145941734, + 0.011066222563385963, + 0.006750849075615406, + 0.004460908938199282, + 0.008172848261892796, + 0.005305177997797728, + 0.03523140028119087, + 0.01642884686589241, + 0.23497803509235382, + 0.066108837723732, + 0.022812742739915848, + 0.009668678976595402, + 0.048895832151174545, + 0.08454718440771103, + 0.003379916539415717, + 0.005034063011407852, + 0.0036024623550474644, + 0.004983983933925629, + 0.006911145057529211, + 0.004893920850008726, + 0.003619241528213024, + 0.004296339116990566, + 0.002220835303887725, + 0.00808953307569027, + 0.007496223784983158, + 0.0034176763147115707, + 0.004117083270102739, + 0.002580903470516205, + 0.0027002724818885326, + 0.0024447094183415174, + 0.0028232985641807318, + 0.0030508609488606453, + 0.0072390190325677395, + 0.005929549224674702, + 0.002726291073486209, + 0.005091961473226547, + 0.0040945978835225105, + 0.008342226035892963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.17548047006130219 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007645629812031984 + ], + "opposite_logprob_tokens": [ + -4.882645606994629 + ], + "logprob_reward": -0.007645629812031984, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007645629812031984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12799911201000214, + 0.039038848131895065, + 0.08708582818508148, + 0.028629666194319725, + 0.02691168710589409, + 0.02698548138141632, + 0.40116968750953674, + 0.021268144249916077, + 0.13980022072792053, + 0.10111130028963089 + ], + "credit_map": [ + 0.01073545217514038, + 0.026698987931013107, + 0.02084580808877945, + 0.01175652164965868, + 0.028881365433335304, + 0.011658323928713799, + 0.008204812183976173, + 0.009217832237482071, + 0.003701288253068924, + 0.004111479036509991, + 0.004610288422554731, + 0.006461013574153185, + 0.008260966278612614, + 0.0026155170053243637, + 0.003976757172495127, + 0.0053015341982245445, + 0.007426099386066198, + 0.010588749311864376, + 0.006844943854957819, + 0.014022614806890488, + 0.02037280797958374, + 0.008209708146750927, + 0.011949595995247364, + 0.007671298924833536, + 0.0022387951612472534, + 0.0025140708312392235, + 0.003480680054053664, + 0.006780228577554226, + 0.00288851885125041, + 0.0019584945403039455, + 0.006033305544406176, + 0.0027355686761438847, + 0.0018924471223726869, + 0.002120209624990821, + 0.004614597652107477, + 0.0051886155270040035, + 0.0028007933869957924, + 0.0032054483890533447, + 0.004460541531443596, + 0.0026290311943739653, + 0.003777120728045702, + 0.005623613018542528, + 0.004388729576021433, + 0.004554571118205786, + 0.0015947036445140839, + 0.001967662712559104, + 0.003055472858250141, + 0.002023605164140463, + 0.028428370133042336, + 0.014635046012699604, + 0.18164139986038208, + 0.049670714884996414, + 0.017579054459929466, + 0.00801900215446949, + 0.036021701991558075, + 0.06517437100410461, + 0.0016626444412395358, + 0.0024998104199767113, + 0.004094310570508242, + 0.0028787816409021616, + 0.0030726499389857054, + 0.002682809019461274, + 0.0022583166137337685, + 0.0021188203245401382, + 0.028032032772898674, + 0.05714363604784012, + 0.008543542586266994, + 0.006537500768899918, + 0.018557725474238396, + 0.009813067503273487, + 0.004583874251693487, + 0.0065888226963579655, + 0.007090744096785784, + 0.011594696901738644, + 0.01595696620643139, + 0.012396479956805706, + 0.007060742471367121, + 0.019210267812013626, + 0.010349936783313751, + 0.0174514539539814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.007645629812031984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871440887451172 + ], + "opposite_logprob_tokens": [ + -1.1371440887451172 + ], + "logprob_reward": -0.3871440887451172, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871440887451172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10672949254512787, + 0.033010829240083694, + 0.18271397054195404, + 0.03066794015467167, + 0.046538084745407104, + 0.0404072031378746, + 0.4165121018886566, + 0.0343092679977417, + 0.04730988293886185, + 0.061801232397556305 + ], + "credit_map": [ + 0.010893735103309155, + 0.01018297765403986, + 0.018175452947616577, + 0.01854666694998741, + 0.009794030338525772, + 0.008164196275174618, + 0.018481584265828133, + 0.012490853667259216, + 0.003238425822928548, + 0.003584266873076558, + 0.003981985617429018, + 0.006762910634279251, + 0.005504850298166275, + 0.0024970348458737135, + 0.0038800621405243874, + 0.0035612958017736673, + 0.015657953917980194, + 0.024510947987437248, + 0.013586322776973248, + 0.018992604687809944, + 0.044385917484760284, + 0.016476646065711975, + 0.030006036162376404, + 0.019097553566098213, + 0.0025360090658068657, + 0.002349396701902151, + 0.003561796387657523, + 0.006694645620882511, + 0.004389574751257896, + 0.002928841160610318, + 0.005722872447222471, + 0.002484806813299656, + 0.0029531472828239202, + 0.0034259562380611897, + 0.00817921757698059, + 0.009137511253356934, + 0.003911152947694063, + 0.004403548780828714, + 0.008731292560696602, + 0.00579626252874732, + 0.005709259770810604, + 0.0035660734865814447, + 0.008184568956494331, + 0.004713216330856085, + 0.003246246837079525, + 0.0034578684717416763, + 0.007985042408108711, + 0.003544932696968317, + 0.024345995858311653, + 0.010580956004559994, + 0.19311614334583282, + 0.05177660658955574, + 0.017456067726016045, + 0.010910924524068832, + 0.038624465465545654, + 0.0697009339928627, + 0.002519527915865183, + 0.0041059707291424274, + 0.006110718939453363, + 0.004510176833719015, + 0.005032395478338003, + 0.004958059173077345, + 0.003225539578124881, + 0.0038468812126666307, + 0.008613322861492634, + 0.01878497563302517, + 0.0024690122809261084, + 0.002787611912935972, + 0.0068929181434214115, + 0.0034842786844819784, + 0.002384650520980358, + 0.001893115695565939, + 0.004935338627547026, + 0.008660486899316311, + 0.010187213309109211, + 0.008621951565146446, + 0.005313670728355646, + 0.01239291112869978, + 0.005562126636505127, + 0.00612753676250577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.3871440887451172 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086530983448029 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086530983448029, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086530983448029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12093166261911392, + 0.04013317823410034, + 0.2493433654308319, + 0.03491748869419098, + 0.05388620123267174, + 0.060721490532159805, + 0.19080467522144318, + 0.04367996007204056, + 0.13459758460521698, + 0.07098442316055298 + ], + "credit_map": [ + 0.013165882788598537, + 0.011188133619725704, + 0.02091955579817295, + 0.023425910621881485, + 0.010735110379755497, + 0.010882758535444736, + 0.016413893550634384, + 0.014200415462255478, + 0.004048088565468788, + 0.0050346157513558865, + 0.004077007528394461, + 0.00722176767885685, + 0.006132895592600107, + 0.002502249786630273, + 0.0043014781549572945, + 0.006815072614699602, + 0.02109561674296856, + 0.033382076770067215, + 0.017153868451714516, + 0.02411053515970707, + 0.06146905943751335, + 0.02199098840355873, + 0.04201342537999153, + 0.028127795085310936, + 0.0028711764607578516, + 0.0032276560086756945, + 0.0057512009516358376, + 0.007265792693942785, + 0.0035483422689139843, + 0.0030920719727873802, + 0.006391502916812897, + 0.0027697421610355377, + 0.003796470584347844, + 0.004384365864098072, + 0.008403770625591278, + 0.013305321335792542, + 0.004450722597539425, + 0.0042955875396728516, + 0.008264811709523201, + 0.006985146552324295, + 0.009190698154270649, + 0.008714417926967144, + 0.009876205585896969, + 0.008906164206564426, + 0.005102840252220631, + 0.005803578067570925, + 0.008252600207924843, + 0.004874980077147484, + 0.012961086817085743, + 0.011783775873482227, + 0.07634316384792328, + 0.02082803100347519, + 0.013459320180118084, + 0.008792426437139511, + 0.01642598770558834, + 0.030210861936211586, + 0.003248253371566534, + 0.004930910188704729, + 0.007893959991633892, + 0.005831290036439896, + 0.0061660632491111755, + 0.00626795319840312, + 0.004344838671386242, + 0.004996688570827246, + 0.02618025802075863, + 0.056083012372255325, + 0.0063334740698337555, + 0.006724203005433083, + 0.01924583874642849, + 0.009401967748999596, + 0.005231543909758329, + 0.005397275555878878, + 0.006120243575423956, + 0.0087991151958704, + 0.012293714098632336, + 0.010268216021358967, + 0.006387089379131794, + 0.011994794942438602, + 0.0068607511930167675, + 0.00826049130409956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.011086530983448029 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097782202064991 + ], + "opposite_logprob_tokens": [ + -4.634778022766113 + ], + "logprob_reward": -0.0097782202064991, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097782202064991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17094960808753967, + 0.06374899297952652, + 0.04020128771662712, + 0.034525543451309204, + 0.07276197522878647, + 0.06793474406003952, + 0.3824421763420105, + 0.020530376583337784, + 0.08034411817789078, + 0.06656121462583542 + ], + "credit_map": [ + 0.010961472988128662, + 0.028739506378769875, + 0.03666263818740845, + 0.03204665333032608, + 0.025341657921671867, + 0.009176286868751049, + 0.016317211091518402, + 0.011704164557158947, + 0.005877357441931963, + 0.006964317988604307, + 0.007796857506036758, + 0.011456136591732502, + 0.01219151634722948, + 0.00484533840790391, + 0.006940023973584175, + 0.0076774419285357, + 0.003265887266024947, + 0.004201917443424463, + 0.004869917873293161, + 0.008408022113144398, + 0.00726786395534873, + 0.002361919730901718, + 0.00525559950619936, + 0.004570163320749998, + 0.0024913386441767216, + 0.002879706909880042, + 0.0038726755883544683, + 0.008895649574697018, + 0.0054986621253192425, + 0.0024058313574641943, + 0.004788909573107958, + 0.0036927699111402035, + 0.006248416844755411, + 0.008942636661231518, + 0.008125517517328262, + 0.00885450653731823, + 0.013504465110599995, + 0.009123335592448711, + 0.007349400781095028, + 0.010613693855702877, + 0.005493976641446352, + 0.007969396188855171, + 0.015253368765115738, + 0.010080114006996155, + 0.007819385267794132, + 0.0069244857877492905, + 0.008135819807648659, + 0.006258200388401747, + 0.02555139735341072, + 0.027806825935840607, + 0.14470990002155304, + 0.0446881502866745, + 0.02526978775858879, + 0.013847497291862965, + 0.03592003136873245, + 0.0646485984325409, + 0.002189702820032835, + 0.005108451470732689, + 0.0018613830907270312, + 0.0017616933910176158, + 0.0035755913704633713, + 0.002431616885587573, + 0.0016998662613332272, + 0.0019020700128749013, + 0.01505438145250082, + 0.025124192237854004, + 0.012311110273003578, + 0.005454729776829481, + 0.008055075071752071, + 0.006063009146600962, + 0.0037722985725849867, + 0.004509316757321358, + 0.004996192175894976, + 0.0069168987683951855, + 0.010986152105033398, + 0.008930426090955734, + 0.00467498367652297, + 0.010525228455662727, + 0.007238141726702452, + 0.012293187901377678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.0097782202064991 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An initially static, extreme wide shot slowly tilts right, tracking a mesmerizing natural event across a vast meadow bathed in the warm, golden afternoon sun. A dense, shimmering cloud of hawthorn seeds drifts effortlessly on the breeze. Each seed, a tiny black achene, hovers atop a delicate, fluffy white nipple, catching the low-angle sunlight, causing the entire swarm to emit an ethereal, backlit glow. They soar, turn, and tumble in a slow, chaotic, yet graceful dance, some rising high against the soft blue sky while others skim the tips of tall green grass and scattered wildflowers below. The vast field stretches to the hazy, distant treeline, where long, soft shadows cast by the setting sun accentuate the texture of the landscape. The overall atmosphere is one of profound tranquility and dreamlike wonder, capturing a fleeting, magical moment in nature.", + "seed": 19830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_5/index_5615_seed_19830_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11419460568577051, + "vlm_logprob": -0.11419460568577051, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.11419460568577051, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21453219652175903 + ], + "opposite_logprob_tokens": [ + -1.7145321369171143 + ], + "logprob_reward": -0.21453219652175903, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21453219652175903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0773167535662651, + 0.04174419119954109, + 0.01894880272448063, + 0.3651452660560608, + 0.10424944758415222, + 0.04991649091243744, + 0.0560222789645195, + 0.11587067693471909, + 0.12555736303329468, + 0.04522868990898132 + ], + "credit_map": [ + 0.01066808495670557, + 0.012837491929531097, + 0.010687786154448986, + 0.009858809411525726, + 0.0049385507591068745, + 0.00558043085038662, + 0.009084357880055904, + 0.013661239296197891, + 0.003857637755572796, + 0.0050962818786501884, + 0.007743649650365114, + 0.006440239492803812, + 0.003079906338825822, + 0.0028865570202469826, + 0.004542679060250521, + 0.008097237907350063, + 0.0016298580449074507, + 0.0025276823434978724, + 0.002389177680015564, + 0.0022459817118942738, + 0.0018053982639685273, + 0.001240192330442369, + 0.0028658336959779263, + 0.004244678653776646, + 0.0471489317715168, + 0.11664911359548569, + 0.017890257760882378, + 0.01609828881919384, + 0.018839912489056587, + 0.014621296897530556, + 0.041903767734766006, + 0.09199368953704834, + 0.00790178682655096, + 0.022551707923412323, + 0.011446913704276085, + 0.020628001540899277, + 0.009505088441073895, + 0.005990524310618639, + 0.014464256353676319, + 0.011761164292693138, + 0.0033742431551218033, + 0.006198100280016661, + 0.0056418366730213165, + 0.005918301176279783, + 0.003320803167298436, + 0.0016812434187158942, + 0.012044215574860573, + 0.011737748049199581, + 0.004860223736613989, + 0.006883175112307072, + 0.006227347534149885, + 0.006330310832709074, + 0.005382878705859184, + 0.003790120128542185, + 0.014327679760754108, + 0.008220539428293705, + 0.022897083312273026, + 0.009127587080001831, + 0.007359766401350498, + 0.010714659467339516, + 0.009755171835422516, + 0.006960163824260235, + 0.016523178666830063, + 0.03253306820988655, + 0.017291471362113953, + 0.007710699457675219, + 0.012408032082021236, + 0.026529403403401375, + 0.010772617533802986, + 0.00573005760088563, + 0.023244809359312057, + 0.02187027409672737, + 0.003142985515296459, + 0.003107713768258691, + 0.00641080504283309, + 0.007168426178395748, + 0.0032277649734169245, + 0.0036612257827073336, + 0.012991807423532009, + 0.005517958197742701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.21453219652175903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601341500878334 + ], + "opposite_logprob_tokens": [ + -3.895601272583008 + ], + "logprob_reward": -0.020601341500878334, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601341500878334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1794467270374298, + 0.10710693150758743, + 0.04117097705602646, + 0.30956199765205383, + 0.08281739056110382, + 0.06820107251405716, + 0.028602510690689087, + 0.038581766188144684, + 0.09890351444482803, + 0.045607101172208786 + ], + "credit_map": [ + 0.02433617040514946, + 0.04702654853463173, + 0.02108789049088955, + 0.02191237173974514, + 0.012119513005018234, + 0.012874550186097622, + 0.01809464953839779, + 0.02199503220617771, + 0.015412546694278717, + 0.02432604320347309, + 0.012793343514204025, + 0.012737923301756382, + 0.009766025468707085, + 0.006990812718868256, + 0.008668129332363605, + 0.01641211099922657, + 0.0031542545184493065, + 0.006306298077106476, + 0.004129024222493172, + 0.00749038765206933, + 0.002664032857865095, + 0.002279627835378051, + 0.005618033464998007, + 0.009529318660497665, + 0.03252516686916351, + 0.08659107238054276, + 0.01836797036230564, + 0.015431676991283894, + 0.017723267897963524, + 0.012397082522511482, + 0.04297835752367973, + 0.0835474282503128, + 0.0067710066214203835, + 0.014052892103791237, + 0.008508668281137943, + 0.016701070591807365, + 0.006573875900357962, + 0.004803885240107775, + 0.010854353196918964, + 0.01455164235085249, + 0.005941336043179035, + 0.009232264943420887, + 0.011879962868988514, + 0.012278796173632145, + 0.004088300280272961, + 0.0041158986277878284, + 0.012796616181731224, + 0.007867900654673576, + 0.0026356822345405817, + 0.002198465634137392, + 0.003163045272231102, + 0.004060179926455021, + 0.002017480554059148, + 0.0017213240498676896, + 0.005102539900690317, + 0.007703793700784445, + 0.007466863375157118, + 0.0031714672222733498, + 0.0033261796925216913, + 0.0039630974642932415, + 0.0036558962892740965, + 0.002656713593751192, + 0.004648412112146616, + 0.009693131782114506, + 0.01377861388027668, + 0.0041825738735497, + 0.0054511819034814835, + 0.008340292610228062, + 0.0034275942016392946, + 0.003361235372722149, + 0.025633113458752632, + 0.03472890332341194, + 0.003358672372996807, + 0.0026709865778684616, + 0.006589023396372795, + 0.006869371980428696, + 0.002702240366488695, + 0.0037651918828487396, + 0.013735661283135414, + 0.005915955640375614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.020601341500878334 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135196566581726 + ], + "opposite_logprob_tokens": [ + -1.3135197162628174 + ], + "logprob_reward": -0.3135196566581726, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135196566581726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22220157086849213, + 0.04633112624287605, + 0.0566423162817955, + 0.1210370883345604, + 0.13812117278575897, + 0.09093860536813736, + 0.056536369025707245, + 0.028812602162361145, + 0.12656952440738678, + 0.11280960589647293 + ], + "credit_map": [ + 0.034276269376277924, + 0.07657237350940704, + 0.01965462788939476, + 0.02169664390385151, + 0.013704290613532066, + 0.014097166247665882, + 0.018260430544614792, + 0.02393976040184498, + 0.006560636684298515, + 0.010415015742182732, + 0.0044270046055316925, + 0.00540917320176959, + 0.0038000699132680893, + 0.002655615098774433, + 0.00471832137554884, + 0.008345289155840874, + 0.004839185159653425, + 0.008797269314527512, + 0.006121819838881493, + 0.008352750912308693, + 0.00310856313444674, + 0.002484144875779748, + 0.008585692383348942, + 0.014352889731526375, + 0.01383677776902914, + 0.032521020621061325, + 0.010069238021969795, + 0.008841694332659245, + 0.006420108489692211, + 0.005475636571645737, + 0.01771288551390171, + 0.026159731671214104, + 0.012861030176281929, + 0.019098125398159027, + 0.015835190191864967, + 0.028930900618433952, + 0.011219875887036324, + 0.008859478868544102, + 0.01851641573011875, + 0.0228001456707716, + 0.006000684108585119, + 0.012649989686906338, + 0.017032507807016373, + 0.017468707635998726, + 0.006124288309365511, + 0.003983499947935343, + 0.015422881580889225, + 0.012256045825779438, + 0.004144147504121065, + 0.002734534675255418, + 0.005218457896262407, + 0.006197766866534948, + 0.005137498024851084, + 0.0036025766748934984, + 0.013726774603128433, + 0.015774616971611977, + 0.002477629343047738, + 0.0018558480078354478, + 0.004142062272876501, + 0.005197037477046251, + 0.002740489086136222, + 0.0020606538746505976, + 0.005747516639530659, + 0.004591364413499832, + 0.016258103772997856, + 0.007405453361570835, + 0.009091120213270187, + 0.019769269973039627, + 0.005039198789745569, + 0.005116017535328865, + 0.030608689412474632, + 0.03328166529536247, + 0.007670613005757332, + 0.007211007177829742, + 0.017558004707098007, + 0.017923889681696892, + 0.008214750327169895, + 0.008273991756141186, + 0.03264942392706871, + 0.013307922519743443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.3135196566581726 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785657748579979 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785657748579979, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785657748579979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17135314643383026, + 0.05145691707730293, + 0.03947791829705238, + 0.32207146286964417, + 0.12761715054512024, + 0.06970376521348953, + 0.045272041112184525, + 0.029709678143262863, + 0.06056657433509827, + 0.08277136832475662 + ], + "credit_map": [ + 0.02556671015918255, + 0.05037620663642883, + 0.020636824890971184, + 0.019051460549235344, + 0.010147719644010067, + 0.009699209593236446, + 0.01748468726873398, + 0.018390333279967308, + 0.007641437463462353, + 0.012880661524832249, + 0.004529816098511219, + 0.006035161670297384, + 0.0046628545969724655, + 0.00339123560115695, + 0.004276329185813665, + 0.008039417676627636, + 0.0029472599271684885, + 0.006540543399751186, + 0.004608331713825464, + 0.006537305656820536, + 0.002506064949557185, + 0.0021137073636054993, + 0.005298098083585501, + 0.00892660953104496, + 0.0394185408949852, + 0.09843795746564865, + 0.016902336850762367, + 0.014889044687151909, + 0.015608379617333412, + 0.012553278356790543, + 0.04097766429185867, + 0.08328425139188766, + 0.010427710600197315, + 0.02144371159374714, + 0.014353190548717976, + 0.02684931270778179, + 0.010653751902282238, + 0.007645559497177601, + 0.017071887850761414, + 0.019172031432390213, + 0.004599283915013075, + 0.009680468589067459, + 0.013189279474318027, + 0.01339873019605875, + 0.004728071391582489, + 0.0032265842892229557, + 0.011349056847393513, + 0.009532292373478413, + 0.003326451638713479, + 0.0026676137931644917, + 0.004801867064088583, + 0.0053268619813025, + 0.004186686128377914, + 0.003723763395100832, + 0.011341586709022522, + 0.009897206909954548, + 0.0036276732571423054, + 0.0023706911597400904, + 0.0033128922805190086, + 0.00460899667814374, + 0.002571003045886755, + 0.0020675852429121733, + 0.005286697298288345, + 0.005864137317985296, + 0.006676170509308577, + 0.004292081110179424, + 0.004961577709764242, + 0.010919295251369476, + 0.0027384315617382526, + 0.003416482824832201, + 0.014596423134207726, + 0.012966112233698368, + 0.0056539978832006454, + 0.005326559767127037, + 0.013035908341407776, + 0.012361244298517704, + 0.005709055345505476, + 0.006145250052213669, + 0.024773791432380676, + 0.009765556082129478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.009785657748579979 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an extreme wide shot that slowly tilts right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a vast meadow bathed in warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are there tiny black achenes with delicate, fluffy white nipples drifting in a dense cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the seeds exhibit slow, chaotic, yet graceful motion including soaring, turning, and tumbling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft blue sky with a hazy distant treeline and long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534175999462605 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012534175999462605, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534175999462605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07608486711978912, + 0.034710466861724854, + 0.03543851897120476, + 0.7084121704101562, + 0.01832890324294567, + 0.013214810751378536, + 0.012828712351620197, + 0.020022617653012276, + 0.06332819163799286, + 0.017630700021982193 + ], + "credit_map": [ + 0.0074256728403270245, + 0.01483689621090889, + 0.011694507673382759, + 0.011870939284563065, + 0.004772202577441931, + 0.004984417464584112, + 0.00875428318977356, + 0.011745951138436794, + 0.004341396503150463, + 0.0077680665999650955, + 0.0047973706386983395, + 0.004192556720227003, + 0.00300517282448709, + 0.002296976512297988, + 0.0030907588079571724, + 0.005218168254941702, + 0.0024462915025651455, + 0.005043238867074251, + 0.003416866296902299, + 0.006729212589561939, + 0.0025699245743453503, + 0.002268501091748476, + 0.005186317954212427, + 0.007778168190270662, + 0.0826883614063263, + 0.21514691412448883, + 0.03767412155866623, + 0.03178319334983826, + 0.038367096334695816, + 0.02690548449754715, + 0.08965086191892624, + 0.18619617819786072, + 0.001975767081603408, + 0.003507329151034355, + 0.0016769226640462875, + 0.0016096162144094706, + 0.001575515023432672, + 0.0009308496955782175, + 0.002785791177302599, + 0.004267111420631409, + 0.0015475749969482422, + 0.001401541638188064, + 0.0020353880245238543, + 0.00199390877969563, + 0.0011219201842322946, + 0.0006452841334976256, + 0.0019910254050046206, + 0.0024781671818345785, + 0.0013195815263316035, + 0.0012733028270304203, + 0.0013975745532661676, + 0.002093480434268713, + 0.0017262301407754421, + 0.001307566068135202, + 0.002034268341958523, + 0.0016767082270234823, + 0.0025553444866091013, + 0.0024020590353757143, + 0.0017262522596865892, + 0.0017851968295872211, + 0.0027893963269889355, + 0.0022975895553827286, + 0.003386355470865965, + 0.0030804243870079517, + 0.005000591278076172, + 0.0024283293168991804, + 0.003612673841416836, + 0.00822119414806366, + 0.0022871936671435833, + 0.002006270457059145, + 0.02021356113255024, + 0.019558381289243698, + 0.0015181803610175848, + 0.0017644613981246948, + 0.0022007059305906296, + 0.0022326826583594084, + 0.0020087435841560364, + 0.0021517956629395485, + 0.0035785057116299868, + 0.002175623085349798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.012534175999462605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_6/index_tensor([7596, 7596])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_6/index_tensor([7596, 7596])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..63f1bd1372f52c8d2b234e9430a49f9fbe6663b1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_6/index_tensor([7596, 7596])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle shot, slowly and steadily moving forward, focuses on a corner of a dark, solid wood desk. A soft, yet somewhat cold, afternoon light shone from a side window onto the desk. A printed diagnosis report lay neatly spread out, its glaring white contrasting with the dullness of the surface. A pair of unused reading glasses sat beside it. An elderly hand, its skin wrinkled and lightly freckled, entered the frame, trembling slightly but steadily with uncontrollable nervousness. Its index finger, carefully touching the cold paper, moved with extreme slowness and hesitation along the lines of black print, as if reading each word of an unacceptable verdict. As the camera zoomed in, we could clearly see every wrinkle in the finger's skin, the veins tense with tension. Finally, at the end of a crucial line containing the diagnosis, the trembling finger paused, frozen in a resolute and powerless motion. The entire scene, shrouded in silence, was permeated with a heavy, oppressive sense of fate.", + "seed": 25675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_6/index_7596_seed_25675_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.038478844612836835, + "vlm_logprob": -0.038478844612836835, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.038478844612836835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02680356241762638 + ], + "opposite_logprob_tokens": [ + -4.776803493499756 + ], + "logprob_reward": -0.02680356241762638, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.02680356241762638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40124377608299255, + 0.142306387424469, + 0.05827280506491661, + 0.05741750821471214, + 0.05068722367286682, + 0.06454767286777496, + 0.06467650830745697, + 0.05635638162493706, + 0.05194978043437004, + 0.052541863173246384 + ], + "credit_map": [ + 0.02636360190808773, + 0.02910918928682804, + 0.061686452478170395, + 0.14293839037418365, + 0.06944733113050461, + 0.039337050169706345, + 0.018998917192220688, + 0.013362823985517025, + 0.013949703425168991, + 0.017031284049153328, + 0.03304702788591385, + 0.020579025149345398, + 0.014669223688542843, + 0.013633291237056255, + 0.008884521201252937, + 0.020512321963906288, + 0.00803825631737709, + 0.006122016813606024, + 0.01068048644810915, + 0.007262481842190027, + 0.004302063025534153, + 0.005904768593609333, + 0.007765330374240875, + 0.008197406306862831, + 0.01004093699157238, + 0.006253163330256939, + 0.01177265029400587, + 0.0068449582904577255, + 0.005101304035633802, + 0.0058882469311356544, + 0.005338035523891449, + 0.006178215146064758, + 0.006557701621204615, + 0.006482536438852549, + 0.00814000703394413, + 0.009020859375596046, + 0.004273569211363792, + 0.004663895815610886, + 0.006781728006899357, + 0.004766924772411585, + 0.0069662658497691154, + 0.007404169533401728, + 0.011568328365683556, + 0.011130212806165218, + 0.007731067016720772, + 0.008479932323098183, + 0.006955553311854601, + 0.004312149714678526, + 0.005393845494836569, + 0.008423002436757088, + 0.006148624233901501, + 0.011048895306885242, + 0.006315518170595169, + 0.01064744871109724, + 0.00835882592946291, + 0.008340348489582539, + 0.00549060245975852, + 0.006548942066729069, + 0.008681872859597206, + 0.007559145335108042, + 0.007885727100074291, + 0.007778416853398085, + 0.007536022458225489, + 0.004875654820352793, + 0.006568257696926594, + 0.004665099084377289, + 0.006062426138669252, + 0.005350199993699789, + 0.004907725844532251, + 0.009532676078379154, + 0.008398857899010181, + 0.00646453769877553, + 0.005077534355223179, + 0.0038257099222391844, + 0.004765819758176804, + 0.004680551588535309, + 0.006594369653612375, + 0.012159002013504505, + 0.00863715074956417, + 0.006801729090511799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02680356241762638 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812900349497795 + ], + "opposite_logprob_tokens": [ + -3.529812812805176 + ], + "logprob_reward": -0.029812900349497795, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812900349497795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3135181963443756, + 0.06592933088541031, + 0.02952159196138382, + 0.11025355011224747, + 0.12014611065387726, + 0.255408376455307, + 0.03503461927175522, + 0.02211148291826248, + 0.016870057210326195, + 0.031206708401441574 + ], + "credit_map": [ + 0.018489092588424683, + 0.026566024869680405, + 0.048807159066200256, + 0.11084318161010742, + 0.04646935313940048, + 0.03147156164050102, + 0.017089199274778366, + 0.013782620429992676, + 0.0077301557175815105, + 0.008432254195213318, + 0.01339677069336176, + 0.010260180570185184, + 0.004705366678535938, + 0.0061824931763112545, + 0.008867118507623672, + 0.006354998331516981, + 0.003859046846628189, + 0.0030116387642920017, + 0.005614263471215963, + 0.0038874659221619368, + 0.0017794041195884347, + 0.0026681111194193363, + 0.004479151219129562, + 0.004222510382533073, + 0.008662141859531403, + 0.014452225528657436, + 0.013402112759649754, + 0.0209579449146986, + 0.008837270550429821, + 0.015416360460221767, + 0.015032630413770676, + 0.013492865487933159, + 0.017495520412921906, + 0.019477128982543945, + 0.01804848574101925, + 0.012785464525222778, + 0.013887754641473293, + 0.011276543140411377, + 0.012007545679807663, + 0.015167677775025368, + 0.021761490032076836, + 0.024186694994568825, + 0.039944518357515335, + 0.046751804649829865, + 0.03086771070957184, + 0.03208973631262779, + 0.03688208758831024, + 0.022924348711967468, + 0.003240464022383094, + 0.005570379551500082, + 0.004410949535667896, + 0.0063715483993291855, + 0.004417191259562969, + 0.005055266432464123, + 0.003703697817400098, + 0.002265121089294553, + 0.002794082975015044, + 0.00152452546171844, + 0.0031145811080932617, + 0.0024960609152913094, + 0.0024259972851723433, + 0.004272195510566235, + 0.0031275558285415173, + 0.0023564836010336876, + 0.001936459681019187, + 0.0015399126568809152, + 0.001879538525827229, + 0.0023535604123026133, + 0.001787162502296269, + 0.0032939231023192406, + 0.002537214895710349, + 0.0015422871802002192, + 0.0036080635618418455, + 0.00384926307015121, + 0.003436531638726592, + 0.004377248231321573, + 0.003361242823302746, + 0.005723058246076107, + 0.00412044208496809, + 0.002730858977884054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.029812900349497795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904840797185898 + ], + "opposite_logprob_tokens": [ + -2.464048385620117 + ], + "logprob_reward": -0.08904840797185898, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904840797185898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37586265802383423, + 0.2345159500837326, + 0.016435228288173676, + 0.060123998671770096, + 0.0709267184138298, + 0.14461268484592438, + 0.02744935266673565, + 0.02220250852406025, + 0.021909013390541077, + 0.025961868464946747 + ], + "credit_map": [ + 0.02771768346428871, + 0.046189289540052414, + 0.06966114789247513, + 0.10289730876684189, + 0.04371465742588043, + 0.05039334297180176, + 0.01883748732507229, + 0.016451753675937653, + 0.02880791760981083, + 0.036884061992168427, + 0.0406026765704155, + 0.033131685107946396, + 0.026691872626543045, + 0.021395208314061165, + 0.022126516327261925, + 0.024875998497009277, + 0.0017787764081731439, + 0.002097995486110449, + 0.00325557473115623, + 0.0022796771954745054, + 0.0014813649468123913, + 0.00192075886297971, + 0.0019142215605825186, + 0.0017068603774532676, + 0.011190427467226982, + 0.00636023934930563, + 0.012123150750994682, + 0.007388274185359478, + 0.004743428900837898, + 0.006226360332220793, + 0.005238398443907499, + 0.006853719241917133, + 0.01070726104080677, + 0.0066525898873806, + 0.01292679738253355, + 0.00804454181343317, + 0.010642671957612038, + 0.006659433711320162, + 0.008876865729689598, + 0.006416556425392628, + 0.018256550654768944, + 0.0195061843842268, + 0.022390633821487427, + 0.017672643065452576, + 0.01776391640305519, + 0.018749503418803215, + 0.020482858642935753, + 0.009790388867259026, + 0.003078932175412774, + 0.0027224610093981028, + 0.003307164879515767, + 0.00240353774279356, + 0.00205477187409997, + 0.005120574031025171, + 0.004064637701958418, + 0.0046972730197012424, + 0.0024410563055425882, + 0.0027070664800703526, + 0.0026186062023043633, + 0.0024638234172016382, + 0.002980670426040888, + 0.003678965149447322, + 0.0028133897576481104, + 0.002498929388821125, + 0.0025978609919548035, + 0.0023414490278810263, + 0.0035506386775523424, + 0.0026224004104733467, + 0.0020889807492494583, + 0.003548713866621256, + 0.0030537471175193787, + 0.002105222549289465, + 0.002695067785680294, + 0.002464021323248744, + 0.0026719896122813225, + 0.0027480535209178925, + 0.003506870474666357, + 0.004843553528189659, + 0.0046975756995379925, + 0.0023347383830696344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.08904840797185898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637331426143646 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637331426143646, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637331426143646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27753832936286926, + 0.04226308315992355, + 0.02757927030324936, + 0.052448820322752, + 0.09791671484708786, + 0.07878705114126205, + 0.0850764587521553, + 0.04705924168229103, + 0.03932468965649605, + 0.25200629234313965 + ], + "credit_map": [ + 0.01810152642428875, + 0.018669387325644493, + 0.038496099412441254, + 0.09949561208486557, + 0.06359445303678513, + 0.018245700746774673, + 0.011267347261309624, + 0.009668163023889065, + 0.005756072234362364, + 0.0071399821899831295, + 0.008118183352053165, + 0.0053694904781877995, + 0.0036411697510629892, + 0.0038918391801416874, + 0.004106613341718912, + 0.004239729605615139, + 0.004605574533343315, + 0.0029898551292717457, + 0.005240499507635832, + 0.004008045420050621, + 0.0017777308821678162, + 0.001971798948943615, + 0.003826129948720336, + 0.0031596366316080093, + 0.004937673918902874, + 0.010785786435008049, + 0.008015264756977558, + 0.006551070604473352, + 0.005734649952501059, + 0.006364365573972464, + 0.0043109082616865635, + 0.005749099422246218, + 0.014413791708648205, + 0.009135589003562927, + 0.01700931042432785, + 0.01187798660248518, + 0.015496025793254375, + 0.012951956130564213, + 0.00994074810296297, + 0.007091303821653128, + 0.010782784782350063, + 0.010741115547716618, + 0.009148665703833103, + 0.011246507987380028, + 0.010080196894705296, + 0.011463604867458344, + 0.009698119945824146, + 0.005626054480671883, + 0.011621538549661636, + 0.013012126088142395, + 0.011580695398151875, + 0.01204380951821804, + 0.009310385212302208, + 0.011753584258258343, + 0.009133773855865002, + 0.006620545405894518, + 0.004949992522597313, + 0.006522782146930695, + 0.0057396735064685345, + 0.006406355649232864, + 0.004890493582934141, + 0.006064044311642647, + 0.00719452416524291, + 0.005291376728564501, + 0.0034522502683103085, + 0.003891737200319767, + 0.005396648775786161, + 0.0051510632038116455, + 0.004527218174189329, + 0.007427482400089502, + 0.004708518739789724, + 0.004769769962877035, + 0.022336769849061966, + 0.022257084026932716, + 0.02217925526201725, + 0.022204235196113586, + 0.03226790577173233, + 0.06338293850421906, + 0.038292452692985535, + 0.029085669666528702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.008637331426143646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809202089905739 + ], + "opposite_logprob_tokens": [ + -3.2880921363830566 + ], + "logprob_reward": -0.03809202089905739, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809202089905739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37539297342300415, + 0.08235913515090942, + 0.047825753688812256, + 0.07766859978437424, + 0.1602081060409546, + 0.07243350893259048, + 0.08274883776903152, + 0.023394450545310974, + 0.012545155361294746, + 0.06542354822158813 + ], + "credit_map": [ + 0.022369084879755974, + 0.025469396263360977, + 0.06568348407745361, + 0.13260973989963531, + 0.06999076902866364, + 0.029668904840946198, + 0.016065308824181557, + 0.013536272570490837, + 0.009453248232603073, + 0.011071898974478245, + 0.01774531416594982, + 0.011348633095622063, + 0.0076122102327644825, + 0.007464382331818342, + 0.005527256056666374, + 0.012136184610426426, + 0.005254145711660385, + 0.004290462005883455, + 0.00865968968719244, + 0.009959999471902847, + 0.003436007769778371, + 0.004011054988950491, + 0.006086001172661781, + 0.006128395441919565, + 0.009729833342134953, + 0.008943850174546242, + 0.015153404325246811, + 0.006853366270661354, + 0.008762731216847897, + 0.012241065502166748, + 0.007304983679205179, + 0.008679366670548916, + 0.018040424212813377, + 0.011099223047494888, + 0.017785126343369484, + 0.019613096490502357, + 0.024433083832263947, + 0.015312159433960915, + 0.03039069101214409, + 0.023534299805760384, + 0.007352812215685844, + 0.010542643256485462, + 0.0103962616994977, + 0.010603011585772038, + 0.009578722529113293, + 0.008347013033926487, + 0.009828926995396614, + 0.005784113425761461, + 0.007039060350507498, + 0.01133534125983715, + 0.008011732250452042, + 0.014451318420469761, + 0.007766487542539835, + 0.012312001548707485, + 0.009837229736149311, + 0.011995659209787846, + 0.0024412223137915134, + 0.002190101658925414, + 0.0031515760347247124, + 0.0028393431566655636, + 0.0030702485237270594, + 0.0038969337474554777, + 0.00367745291441679, + 0.0021275707986205816, + 0.001528287772089243, + 0.0010712668299674988, + 0.001926660886965692, + 0.0017586412141099572, + 0.0013914742739871144, + 0.0019015261204913259, + 0.0018805552972480655, + 0.0010867435485124588, + 0.007357978727668524, + 0.00769828911870718, + 0.007689017802476883, + 0.009298826567828655, + 0.008189845830202103, + 0.010408660396933556, + 0.009249668568372726, + 0.005531259346753359 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.03809202089905739 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, slowly and steadily moving forward, focuses on a corner of a dark, solid wood desk. A soft, yet somewhat cold, afternoon light shone from a side window onto the desk. A printed diagnosis report lay neatly spread out, its glaring white contrasting with the dullness of the surface. A pair of unused reading glasses sat beside it. An elderly hand, its skin wrinkled and lightly freckled, entered the frame, trembling slightly but steadily with uncontrollable nervousness. Its index finger, carefully touching the cold paper, moved with extreme slowness and hesitation along the lines of black print, as if reading each word of an unacceptable verdict. As the camera zoomed in, we could clearly see every wrinkle in the finger's skin, the veins tense with tension. Finally, at the end of a crucial line containing the diagnosis, the trembling finger paused, frozen in a resolute and powerless motion. The entire scene, shrouded in silence, was permeated with a heavy, oppressive sense of fate.", + "seed": 25676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_6/index_7596_seed_25676_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04088266463950276, + "vlm_logprob": -0.04088266463950276, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.04088266463950276, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018704228103160858 + ], + "opposite_logprob_tokens": [ + -4.893704414367676 + ], + "logprob_reward": -0.018704228103160858, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.018704228103160858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2450195848941803, + 0.06141401827335358, + 0.13954468071460724, + 0.0579703226685524, + 0.055930234491825104, + 0.04397911578416824, + 0.17741376161575317, + 0.05851686745882034, + 0.04479024559259415, + 0.11542117595672607 + ], + "credit_map": [ + 0.021028677001595497, + 0.01579245924949646, + 0.0068642268888652325, + 0.01119739469140768, + 0.07041408866643906, + 0.0309066791087389, + 0.07383476942777634, + 0.014981294050812721, + 0.006763222627341747, + 0.0059526097029447556, + 0.004045646172016859, + 0.004085932858288288, + 0.013858831487596035, + 0.008305428549647331, + 0.013604579493403435, + 0.004797765985131264, + 0.015239550732076168, + 0.009272407740354538, + 0.004601167514920235, + 0.008022841066122055, + 0.043461788445711136, + 0.013087457977235317, + 0.03580163046717644, + 0.010057833045721054, + 0.012952336110174656, + 0.011901170015335083, + 0.0035499639343470335, + 0.004194675944745541, + 0.007511841133236885, + 0.006038997322320938, + 0.006731129251420498, + 0.005090208258479834, + 0.008857199922204018, + 0.006144492421299219, + 0.0037003958132117987, + 0.004008949734270573, + 0.004340756684541702, + 0.011501938104629517, + 0.009840874001383781, + 0.007535629905760288, + 0.01001090556383133, + 0.00427057733759284, + 0.0038037339691072702, + 0.003615775378420949, + 0.002331603318452835, + 0.009527478367090225, + 0.0061566210351884365, + 0.0042624217458069324, + 0.022397901862859726, + 0.024907732382416725, + 0.010637663304805756, + 0.02373667061328888, + 0.020264314487576485, + 0.02473803050816059, + 0.022258257493376732, + 0.02847318910062313, + 0.008813202381134033, + 0.013320405967533588, + 0.0025766976177692413, + 0.005098483059555292, + 0.00682139303535223, + 0.007408482953906059, + 0.009339047595858574, + 0.005139156244695187, + 0.006837539840489626, + 0.006411899346858263, + 0.0023019742220640182, + 0.003924290183931589, + 0.004005726892501116, + 0.00846506655216217, + 0.0060181282460689545, + 0.006825621239840984, + 0.01926981285214424, + 0.008141332305967808, + 0.004240722861140966, + 0.007181137800216675, + 0.005686125718057156, + 0.01419888623058796, + 0.01605275459587574, + 0.04065040498971939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.018704228103160858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748621817678213 + ], + "opposite_logprob_tokens": [ + -5.006748676300049 + ], + "logprob_reward": -0.006748621817678213, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748621817678213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4963858723640442, + 0.018692005425691605, + 0.17452339828014374, + 0.04761369153857231, + 0.023177575320005417, + 0.04140261560678482, + 0.0908496081829071, + 0.02879519574344158, + 0.05147521197795868, + 0.027084803208708763 + ], + "credit_map": [ + 0.04142565280199051, + 0.030788857489824295, + 0.01023055985569954, + 0.020587025210261345, + 0.1525135040283203, + 0.05662409961223602, + 0.1605408489704132, + 0.02367536351084709, + 0.002884334186092019, + 0.001954672858119011, + 0.0019613064359873533, + 0.0012122989865019917, + 0.003922301344573498, + 0.0023086145520210266, + 0.002711972687393427, + 0.0017365061212331057, + 0.018598411232233047, + 0.009901546873152256, + 0.0036594162229448557, + 0.010069402866065502, + 0.056257400661706924, + 0.017781022936105728, + 0.04603003337979317, + 0.012226172722876072, + 0.007111052516847849, + 0.006528330501168966, + 0.0030072901863604784, + 0.006739289499819279, + 0.004230710212141275, + 0.004847252741456032, + 0.008479243144392967, + 0.006670527625828981, + 0.003328831633552909, + 0.0034137596376240253, + 0.0017544656293466687, + 0.003131335834041238, + 0.0019630086608231068, + 0.0049947635270655155, + 0.0024376988876610994, + 0.0021537146531045437, + 0.008296343497931957, + 0.0026823878288269043, + 0.0021297463681548834, + 0.0026668168138712645, + 0.0021367857698351145, + 0.008838798850774765, + 0.010177741758525372, + 0.004473999608308077, + 0.01477853488177061, + 0.01616579294204712, + 0.005688926670700312, + 0.010082193650305271, + 0.010080567561089993, + 0.010947255417704582, + 0.009895616210997105, + 0.013210732489824295, + 0.003183043096214533, + 0.0034288971219211817, + 0.0014744974905624986, + 0.003021806012839079, + 0.002827908843755722, + 0.003577128518372774, + 0.0067825536243617535, + 0.0044993627816438675, + 0.009225817397236824, + 0.005596695933490992, + 0.0030573373660445213, + 0.005764017812907696, + 0.004252249840646982, + 0.01201885100454092, + 0.00532538490369916, + 0.006234859582036734, + 0.004174673929810524, + 0.002137766219675541, + 0.0013153643812984228, + 0.0021977273281663656, + 0.0014615480322390795, + 0.003695514053106308, + 0.003427764168009162, + 0.008674447424709797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.006748621817678213 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003134474158287 + ], + "opposite_logprob_tokens": [ + -2.350313425064087 + ], + "logprob_reward": -0.1003134474158287, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003134474158287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5544219017028809, + 0.03564361855387688, + 0.0991726815700531, + 0.02331540174782276, + 0.027372006326913834, + 0.033648956567049026, + 0.1155230924487114, + 0.03150872141122818, + 0.058250248432159424, + 0.02114337868988514 + ], + "credit_map": [ + 0.046345897018909454, + 0.03606504574418068, + 0.01200458686798811, + 0.023195749148726463, + 0.1667817234992981, + 0.06372682750225067, + 0.17923715710639954, + 0.027064912021160126, + 0.003853377653285861, + 0.004089160822331905, + 0.0027382872067391872, + 0.0021627740934491158, + 0.005550905596464872, + 0.004307580180466175, + 0.008793037384748459, + 0.004148496314883232, + 0.011102793738245964, + 0.0085050780326128, + 0.003542959224432707, + 0.005594742484390736, + 0.02674129419028759, + 0.010680783540010452, + 0.024708274751901627, + 0.008296740241348743, + 0.005006010178476572, + 0.0037206027191132307, + 0.00199685781262815, + 0.002127563813701272, + 0.0032254091929644346, + 0.002554122591391206, + 0.002461066236719489, + 0.0022237703669816256, + 0.00352606363594532, + 0.003047630423679948, + 0.0030236125458031893, + 0.003848549211397767, + 0.0022410585079342127, + 0.004108521156013012, + 0.0034185070544481277, + 0.0041580647230148315, + 0.004916640929877758, + 0.0037203095853328705, + 0.0032523220870643854, + 0.002848833566531539, + 0.0027689288835972548, + 0.007757650222629309, + 0.0039937845431268215, + 0.004390489310026169, + 0.019181324169039726, + 0.023104626685380936, + 0.0063096871599555016, + 0.012437710538506508, + 0.0146146509796381, + 0.0120920455083251, + 0.01197761856019497, + 0.015805423259735107, + 0.004727544728666544, + 0.003971698693931103, + 0.0014976445818319917, + 0.003354996209964156, + 0.0038177159149199724, + 0.004820199217647314, + 0.006402625236660242, + 0.002916297409683466, + 0.009831273928284645, + 0.006142441648989916, + 0.002625974826514721, + 0.004745657555758953, + 0.004405722953379154, + 0.010806307196617126, + 0.006579930428415537, + 0.013112939894199371, + 0.003302868688479066, + 0.003005782375112176, + 0.0014994624070823193, + 0.001865471014752984, + 0.0014436037745326757, + 0.004135355353355408, + 0.0021761655807495117, + 0.003714669728651643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.1003134474158287 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863721314817667 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.00863721314817667, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863721314817667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08000031858682632, + 0.024554520845413208, + 0.12425307929515839, + 0.06515195965766907, + 0.05413033068180084, + 0.07046929746866226, + 0.31147530674934387, + 0.1806505173444748, + 0.05637015774846077, + 0.032944533973932266 + ], + "credit_map": [ + 0.008938055485486984, + 0.01131169032305479, + 0.005050504580140114, + 0.004977566655725241, + 0.009993534535169601, + 0.017799012362957, + 0.015013160184025764, + 0.0069167958572506905, + 0.0038122739642858505, + 0.0032122984994202852, + 0.0019484616350382566, + 0.0019742378499358892, + 0.004564287140965462, + 0.003810234135016799, + 0.0028934527654200792, + 0.0023392755538225174, + 0.015809079632163048, + 0.0126469312235713, + 0.006338102743029594, + 0.007524796761572361, + 0.03563180938363075, + 0.0166876632720232, + 0.020432019606232643, + 0.009182668291032314, + 0.013159407302737236, + 0.013529662042856216, + 0.003819909179583192, + 0.00426308810710907, + 0.008556450717151165, + 0.008512450382113457, + 0.00725773349404335, + 0.006053254473954439, + 0.009720423258841038, + 0.005543416831642389, + 0.005134629551321268, + 0.007776608224958181, + 0.004442349076271057, + 0.00956946425139904, + 0.006674397736787796, + 0.005269043147563934, + 0.012194192968308926, + 0.00460053188726306, + 0.003919897135347128, + 0.004841316025704145, + 0.0046324641443789005, + 0.02255295403301716, + 0.011618591845035553, + 0.006109348498284817, + 0.048996903002262115, + 0.05893074721097946, + 0.017099032178521156, + 0.03757331892848015, + 0.038249626755714417, + 0.03219113498926163, + 0.03343786299228668, + 0.04499669373035431, + 0.027005113661289215, + 0.04275863617658615, + 0.00825814064592123, + 0.0129244988784194, + 0.02052866294980049, + 0.023426316678524017, + 0.027284005656838417, + 0.018465131521224976, + 0.009796041063964367, + 0.005025188438594341, + 0.0028775420505553484, + 0.004818914458155632, + 0.0032739313319325447, + 0.007183711510151625, + 0.00644476106390357, + 0.01695007085800171, + 0.0042771571315824986, + 0.0033798657823354006, + 0.0024961179587990046, + 0.004254711326211691, + 0.002279456704854965, + 0.00532118882983923, + 0.00450535211712122, + 0.006430681794881821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.00863721314817667 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000981271266937 + ], + "opposite_logprob_tokens": [ + -2.695009708404541 + ], + "logprob_reward": -0.07000981271266937, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000981271266937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13823837041854858, + 0.07431289553642273, + 0.08976881206035614, + 0.06566863507032394, + 0.09049303084611893, + 0.09892760962247849, + 0.14167897403240204, + 0.049679167568683624, + 0.12111470103263855, + 0.13011780381202698 + ], + "credit_map": [ + 0.014244151301681995, + 0.014978255145251751, + 0.00785130076110363, + 0.009182367473840714, + 0.02681882679462433, + 0.013758540153503418, + 0.03688470646739006, + 0.014520220458507538, + 0.01511419378221035, + 0.008274445310235023, + 0.006356467958539724, + 0.005463087931275368, + 0.008020768873393536, + 0.0082439836114645, + 0.013771199621260166, + 0.009068753570318222, + 0.012337462045252323, + 0.013044624589383602, + 0.0042486111633479595, + 0.009236084297299385, + 0.019101545214653015, + 0.01359676755964756, + 0.010890398174524307, + 0.007313310168683529, + 0.011732109822332859, + 0.009923120029270649, + 0.004778389818966389, + 0.01056419126689434, + 0.0070922765880823135, + 0.008296363987028599, + 0.007394545245915651, + 0.005887641105800867, + 0.01396328303962946, + 0.009089212864637375, + 0.011148495599627495, + 0.012608514167368412, + 0.00791996531188488, + 0.011214394122362137, + 0.014097006060183048, + 0.010452167131006718, + 0.02040618471801281, + 0.009798348881304264, + 0.00779855577275157, + 0.007188487332314253, + 0.007337967399507761, + 0.016531197354197502, + 0.018044646829366684, + 0.011822217144072056, + 0.025990381836891174, + 0.02890762872993946, + 0.009788088500499725, + 0.012660719454288483, + 0.015308786183595657, + 0.015032029710710049, + 0.015710806474089622, + 0.01828054152429104, + 0.00811811350286007, + 0.007630931679159403, + 0.002798793138936162, + 0.005231555551290512, + 0.006643855012953281, + 0.007588624954223633, + 0.006918305531144142, + 0.004748986102640629, + 0.01774653047323227, + 0.008721987716853619, + 0.007379992865025997, + 0.015035601332783699, + 0.006635382771492004, + 0.01860017515718937, + 0.014295423403382301, + 0.03269961103796959, + 0.021807756274938583, + 0.008331059478223324, + 0.005008411128073931, + 0.008845475502312183, + 0.005263770930469036, + 0.017735430970788002, + 0.017235245555639267, + 0.04589066654443741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.07000981271266937 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, slowly and steadily moving forward, focuses on a corner of a dark, solid wood desk. A soft, yet somewhat cold, afternoon light shone from a side window onto the desk. A printed diagnosis report lay neatly spread out, its glaring white contrasting with the dullness of the surface. A pair of unused reading glasses sat beside it. An elderly hand, its skin wrinkled and lightly freckled, entered the frame, trembling slightly but steadily with uncontrollable nervousness. Its index finger, carefully touching the cold paper, moved with extreme slowness and hesitation along the lines of black print, as if reading each word of an unacceptable verdict. As the camera zoomed in, we could clearly see every wrinkle in the finger's skin, the veins tense with tension. Finally, at the end of a crucial line containing the diagnosis, the trembling finger paused, frozen in a resolute and powerless motion. The entire scene, shrouded in silence, was permeated with a heavy, oppressive sense of fate.", + "seed": 25677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_6/index_7596_seed_25677_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06116478014737368, + "vlm_logprob": -0.06116478014737368, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.06116478014737368, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040806736797094345 + ], + "opposite_logprob_tokens": [ + -4.540806770324707 + ], + "logprob_reward": -0.040806736797094345, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.040806736797094345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12321542203426361, + 0.07195308804512024, + 0.0640362948179245, + 0.04805068299174309, + 0.28582000732421875, + 0.05007243901491165, + 0.060542359948158264, + 0.13487961888313293, + 0.07989857345819473, + 0.08153151720762253 + ], + "credit_map": [ + 0.03307495638728142, + 0.008736376650631428, + 0.010058718733489513, + 0.023436134681105614, + 0.008698085322976112, + 0.0077016460709273815, + 0.02180439420044422, + 0.009705102071166039, + 0.007040530908852816, + 0.0052047353237867355, + 0.009173071943223476, + 0.009229017421603203, + 0.010761615820229053, + 0.010176963172852993, + 0.011933664791285992, + 0.00843348354101181, + 0.006573612801730633, + 0.0034311404451727867, + 0.01127183623611927, + 0.008253861218690872, + 0.011007101275026798, + 0.006683580111712217, + 0.007907947525382042, + 0.008907211013138294, + 0.003739151870831847, + 0.0030295720789581537, + 0.00603160634636879, + 0.006314292084425688, + 0.007017502095550299, + 0.005201422143727541, + 0.010943219996988773, + 0.005773912649601698, + 0.026064811274409294, + 0.0207536444067955, + 0.03478014096617699, + 0.028966521844267845, + 0.06789753586053848, + 0.05259108170866966, + 0.02542409673333168, + 0.029342152178287506, + 0.004099524579942226, + 0.007066707592457533, + 0.006406885106116533, + 0.006147775333374739, + 0.007799107581377029, + 0.006502313073724508, + 0.006341592408716679, + 0.005708526354283094, + 0.004742731340229511, + 0.005652802065014839, + 0.007840422913432121, + 0.008445263840258121, + 0.007696741726249456, + 0.00675320765003562, + 0.011683079414069653, + 0.007728107273578644, + 0.009605730883777142, + 0.010359990410506725, + 0.02018212340772152, + 0.01149133313447237, + 0.01740122400224209, + 0.01583225652575493, + 0.01538513321429491, + 0.03462183102965355, + 0.0068603442050516605, + 0.0051874881610274315, + 0.009783021174371243, + 0.008998447097837925, + 0.00857588555663824, + 0.011038483120501041, + 0.007293546572327614, + 0.022161345928907394, + 0.00875487457960844, + 0.006742017809301615, + 0.005753840319812298, + 0.00913126952946186, + 0.013968410901725292, + 0.013212103396654129, + 0.009692191146314144, + 0.014276797883212566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.040806736797094345 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297951638698578 + ], + "opposite_logprob_tokens": [ + -2.2379794120788574 + ], + "logprob_reward": -0.11297951638698578, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297951638698578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03210052102804184, + 0.03903454914689064, + 0.12344995141029358, + 0.26144152879714966, + 0.4647323191165924, + 0.009881352074444294, + 0.0052310326136648655, + 0.03531702607870102, + 0.006170160602778196, + 0.022641586139798164 + ], + "credit_map": [ + 0.003800406353548169, + 0.002719348296523094, + 0.0026267904322594404, + 0.007388615980744362, + 0.002362523227930069, + 0.0032341829501092434, + 0.006018984131515026, + 0.003949668258428574, + 0.0042027318850159645, + 0.0027353332843631506, + 0.006906826514750719, + 0.0054422421380877495, + 0.004797542002052069, + 0.006217076443135738, + 0.005579368211328983, + 0.003153430297970772, + 0.009507820010185242, + 0.011073152534663677, + 0.009314487688243389, + 0.015884650871157646, + 0.026937566697597504, + 0.017062999308109283, + 0.02226424217224121, + 0.0114050367847085, + 0.018611280247569084, + 0.016113851219415665, + 0.026697853580117226, + 0.03633187711238861, + 0.03331105783581734, + 0.029396625235676765, + 0.06410214304924011, + 0.03687683120369911, + 0.04913683608174324, + 0.03550008311867714, + 0.04626056179404259, + 0.049850523471832275, + 0.10530488193035126, + 0.08601810783147812, + 0.0378556102514267, + 0.05480571463704109, + 0.0007218210375867784, + 0.0008083570282906294, + 0.0011666561476886272, + 0.001317404443398118, + 0.0014854574110358953, + 0.0013831055257469416, + 0.001464881468564272, + 0.0015336681390181184, + 0.0005810169386677444, + 0.0005242133047431707, + 0.0004900816711597145, + 0.0008382982923649251, + 0.0006400028360076249, + 0.0006120362086221576, + 0.0007721837027929723, + 0.0007732000085525215, + 0.002163388766348362, + 0.0019077565521001816, + 0.0031868107616901398, + 0.0021030607167631388, + 0.003052160842344165, + 0.007658368442207575, + 0.0037889371160417795, + 0.011456544511020184, + 0.000627820088993758, + 0.000512440805323422, + 0.000923639046959579, + 0.0006754742353223264, + 0.0009784270077943802, + 0.0007434861618094146, + 0.0008286457159556448, + 0.0008802279480732977, + 0.0022965893149375916, + 0.0023030100855976343, + 0.0019402673933655024, + 0.0024163327179849148, + 0.004505389370024204, + 0.0033945452887564898, + 0.0032227342016994953, + 0.0025627193972468376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.11297951638698578 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033468902111053 + ], + "opposite_logprob_tokens": [ + -2.350334644317627 + ], + "logprob_reward": -0.10033468902111053, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033468902111053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07314927875995636, + 0.216610848903656, + 0.08611837029457092, + 0.39153745770454407, + 0.0868319571018219, + 0.016890475526452065, + 0.022869611158967018, + 0.06131080538034439, + 0.013721498660743237, + 0.03095965087413788 + ], + "credit_map": [ + 0.007584753911942244, + 0.006118883844465017, + 0.005799559410661459, + 0.016002308577299118, + 0.005562638398259878, + 0.004905430134385824, + 0.017604384571313858, + 0.009571322239935398, + 0.025883588939905167, + 0.016876855865120888, + 0.036714471876621246, + 0.028369303792715073, + 0.022712528705596924, + 0.029427027329802513, + 0.03363782912492752, + 0.022989239543676376, + 0.006893261801451445, + 0.006802832242101431, + 0.007940675131976604, + 0.012577061541378498, + 0.015774419531226158, + 0.011064067482948303, + 0.015782708302140236, + 0.009283347055315971, + 0.027664050459861755, + 0.024675827473402023, + 0.03936038166284561, + 0.055609457194805145, + 0.04852493479847908, + 0.043087948113679886, + 0.09673532098531723, + 0.05587954446673393, + 0.006544158328324556, + 0.006675794254988432, + 0.006705408450216055, + 0.009241672232747078, + 0.020879829302430153, + 0.016615280881524086, + 0.009800834581255913, + 0.010368977673351765, + 0.0014703775523230433, + 0.0016655660001561046, + 0.002002679742872715, + 0.00217547919601202, + 0.0025074530858546495, + 0.0019813193939626217, + 0.0026381216011941433, + 0.0024494791869074106, + 0.0027115424163639545, + 0.002103108447045088, + 0.0022254472132772207, + 0.0024179713800549507, + 0.003668683348223567, + 0.0025668826419860125, + 0.003695390885695815, + 0.003480584127828479, + 0.004111716989427805, + 0.0027044378221035004, + 0.006792821921408176, + 0.003952774219214916, + 0.004575368948280811, + 0.012516718357801437, + 0.00634022057056427, + 0.020316749811172485, + 0.001358575071208179, + 0.0007884246297180653, + 0.0016934294253587723, + 0.0016885529039427638, + 0.0016241873381659389, + 0.0018526578787714243, + 0.0013751614606007934, + 0.003340510418638587, + 0.0021030788775533438, + 0.0027854933869093657, + 0.0019153036409989, + 0.0023132064379751682, + 0.007648952770978212, + 0.005351479165256023, + 0.003938167821615934, + 0.004903969820588827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.10033468902111053 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642413653433323 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642413653433323, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642413653433323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04036311060190201, + 0.036057233810424805, + 0.04026327654719353, + 0.10814990848302841, + 0.5154001712799072, + 0.030179517343640327, + 0.019718309864401817, + 0.03455140069127083, + 0.015902630984783173, + 0.1594143956899643 + ], + "credit_map": [ + 0.007709020283073187, + 0.002748466795310378, + 0.00416819890961051, + 0.008174783550202847, + 0.003286057384684682, + 0.002901660744100809, + 0.007461191155016422, + 0.003913734573870897, + 0.004305889829993248, + 0.002959158271551132, + 0.005783904809504747, + 0.004778582137078047, + 0.004608651157468557, + 0.005051349755376577, + 0.005412491969764233, + 0.0031572075095027685, + 0.002227124525234103, + 0.001994410064071417, + 0.002245144220069051, + 0.005845729727298021, + 0.007260378450155258, + 0.004691683687269688, + 0.008684097789227962, + 0.007314709480851889, + 0.007477184757590294, + 0.0065537006594240665, + 0.011307824403047562, + 0.014822019264101982, + 0.01455619279295206, + 0.0123439347371459, + 0.02566590905189514, + 0.015423156321048737, + 0.052000176161527634, + 0.03850030526518822, + 0.05734957382082939, + 0.05423517897725105, + 0.11670999228954315, + 0.09412334114313126, + 0.04283898323774338, + 0.05964267626404762, + 0.0020218936260789633, + 0.0032425581011921167, + 0.002600466599687934, + 0.002556305145844817, + 0.007454847916960716, + 0.005252782721072435, + 0.003635193221271038, + 0.003415473271161318, + 0.0014854465844109654, + 0.0019762320443987846, + 0.0018783130217343569, + 0.0017341186758130789, + 0.004304503556340933, + 0.0030394759960472584, + 0.002389127155765891, + 0.0029110945761203766, + 0.002158671384677291, + 0.0012983601773157716, + 0.002480757888406515, + 0.0023580281995236874, + 0.0023140269331634045, + 0.00971077848225832, + 0.0030295727774500847, + 0.011201209388673306, + 0.0014739769976586103, + 0.0015423363074660301, + 0.0019718105904757977, + 0.002303286688402295, + 0.0022614586632698774, + 0.001953307306393981, + 0.002496390137821436, + 0.001900066272355616, + 0.01185510866343975, + 0.014042835682630539, + 0.010322638787329197, + 0.0119593795388937, + 0.019689103588461876, + 0.0307173915207386, + 0.019351132214069366, + 0.04147682711482048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.008642413653433323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430605448782444 + ], + "opposite_logprob_tokens": [ + -3.168060541152954 + ], + "logprob_reward": -0.0430605448782444, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430605448782444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06006626784801483, + 0.04656597971916199, + 0.07416889816522598, + 0.10352760553359985, + 0.1932203620672226, + 0.08654725551605225, + 0.02921127714216709, + 0.04302014037966728, + 0.03348515182733536, + 0.33018702268600464 + ], + "credit_map": [ + 0.010719729587435722, + 0.0069863940589129925, + 0.006946822162717581, + 0.010139001533389091, + 0.0065503790974617004, + 0.004948131274431944, + 0.007889719679951668, + 0.005886094644665718, + 0.0051597305573523045, + 0.0050163655541837215, + 0.006593208760023117, + 0.0050027756951749325, + 0.005235261283814907, + 0.008785140700638294, + 0.0070287929847836494, + 0.0037447039503604174, + 0.0044331662356853485, + 0.003816215554252267, + 0.004659038502722979, + 0.009008186869323254, + 0.019240988418459892, + 0.011714175343513489, + 0.009061931632459164, + 0.012235195375978947, + 0.007957403548061848, + 0.0062906816601753235, + 0.00948523823171854, + 0.015277430415153503, + 0.012943996116518974, + 0.010459070093929768, + 0.026626620441675186, + 0.014487165957689285, + 0.019749531522393227, + 0.012279986403882504, + 0.01695958338677883, + 0.021307434886693954, + 0.043515536934137344, + 0.035922177135944366, + 0.017934096977114677, + 0.025552017614245415, + 0.006383064668625593, + 0.007041614502668381, + 0.009720590896904469, + 0.008555322885513306, + 0.012959645129740238, + 0.010789740830659866, + 0.013572492636740208, + 0.017524784430861473, + 0.0034336617682129145, + 0.003259392222389579, + 0.003237420227378607, + 0.0035008820705115795, + 0.004731707274913788, + 0.003983629867434502, + 0.003443018300458789, + 0.003621567040681839, + 0.0029968349263072014, + 0.002313008299097419, + 0.004786897450685501, + 0.003518023295328021, + 0.0031034161802381277, + 0.01206937525421381, + 0.003394895466044545, + 0.01083768904209137, + 0.002655096584931016, + 0.0029758624732494354, + 0.003653604304417968, + 0.0038295683916658163, + 0.006718390621244907, + 0.004862459842115641, + 0.005099819041788578, + 0.003690354060381651, + 0.02073843590915203, + 0.01453130878508091, + 0.027399778366088867, + 0.025013890117406845, + 0.028297923505306244, + 0.07989100366830826, + 0.026909612119197845, + 0.10740505158901215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.0430605448782444 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, slowly and steadily moving forward, focuses on a corner of a dark, solid wood desk. A soft, yet somewhat cold, afternoon light shone from a side window onto the desk. A printed diagnosis report lay neatly spread out, its glaring white contrasting with the dullness of the surface. A pair of unused reading glasses sat beside it. An elderly hand, its skin wrinkled and lightly freckled, entered the frame, trembling slightly but steadily with uncontrollable nervousness. Its index finger, carefully touching the cold paper, moved with extreme slowness and hesitation along the lines of black print, as if reading each word of an unacceptable verdict. As the camera zoomed in, we could clearly see every wrinkle in the finger's skin, the veins tense with tension. Finally, at the end of a crucial line containing the diagnosis, the trembling finger paused, frozen in a resolute and powerless motion. The entire scene, shrouded in silence, was permeated with a heavy, oppressive sense of fate.", + "seed": 25678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_6/index_7596_seed_25678_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03441062103956938, + "vlm_logprob": -0.03441062103956938, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03441062103956938, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03493383526802063 + ], + "opposite_logprob_tokens": [ + -4.784934043884277 + ], + "logprob_reward": -0.03493383526802063, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.03493383526802063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10918734967708588, + 0.13083676993846893, + 0.04420886188745499, + 0.06727238744497299, + 0.08064546436071396, + 0.14008501172065735, + 0.11704093962907791, + 0.21845830976963043, + 0.05584748461842537, + 0.03641743212938309 + ], + "credit_map": [ + 0.011581480503082275, + 0.0069527099840343, + 0.021372834220528603, + 0.012649661861360073, + 0.009399669244885445, + 0.011309895664453506, + 0.014631176367402077, + 0.021289922297000885, + 0.012687628157436848, + 0.01170726865530014, + 0.01611538603901863, + 0.033171165734529495, + 0.014562693424522877, + 0.010203657671809196, + 0.01551487110555172, + 0.016874101012945175, + 0.005850846879184246, + 0.004915193654596806, + 0.006922517903149128, + 0.004842894151806831, + 0.0027663924265652895, + 0.004908161703497171, + 0.007471639662981033, + 0.006531212944537401, + 0.0030601155012845993, + 0.0036796305794268847, + 0.00813723262399435, + 0.01951347291469574, + 0.014629484154284, + 0.007251095026731491, + 0.005328217055648565, + 0.005673137493431568, + 0.004778535105288029, + 0.008883754722774029, + 0.0069983708672225475, + 0.014733589254319668, + 0.014202705584466457, + 0.01141051109880209, + 0.008902663365006447, + 0.010735335759818554, + 0.008462885394692421, + 0.006775651592761278, + 0.03738420084118843, + 0.019237929955124855, + 0.01476060226559639, + 0.01813819445669651, + 0.008286494761705399, + 0.02703905664384365, + 0.018170693889260292, + 0.01287496741861105, + 0.008649200201034546, + 0.009177343919873238, + 0.024054063484072685, + 0.01699516363441944, + 0.012493671849370003, + 0.014625833369791508, + 0.027873458340764046, + 0.01870977319777012, + 0.02314298041164875, + 0.020733283832669258, + 0.04421549290418625, + 0.03234648331999779, + 0.02565053291618824, + 0.025786299258470535, + 0.006298862397670746, + 0.006208882201462984, + 0.005118068307638168, + 0.005182555411010981, + 0.005708459299057722, + 0.009383541531860828, + 0.006384293083101511, + 0.011562821455299854, + 0.004551817197352648, + 0.0041122715920209885, + 0.00372030446305871, + 0.005242453888058662, + 0.005501140374690294, + 0.005225525237619877, + 0.0036568983923643827, + 0.004407017957419157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 1 + ], + "selected_reward": -0.03493383526802063 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865836814045906 + ], + "opposite_logprob_tokens": [ + -4.758658409118652 + ], + "logprob_reward": -0.00865836814045906, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865836814045906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04963826760649681, + 0.11216704547405243, + 0.07116281241178513, + 0.25904515385627747, + 0.11392593383789062, + 0.20451246201992035, + 0.04528218135237694, + 0.1127888560295105, + 0.01886080764234066, + 0.012616487219929695 + ], + "credit_map": [ + 0.004563856869935989, + 0.0038920945953577757, + 0.004716150928288698, + 0.011605516076087952, + 0.009399844333529472, + 0.004418732598423958, + 0.00504758395254612, + 0.005994480103254318, + 0.009163415990769863, + 0.006979755591601133, + 0.011506973765790462, + 0.02719065546989441, + 0.014417936094105244, + 0.00901083741337061, + 0.01708020269870758, + 0.016817260533571243, + 0.008839897811412811, + 0.005272898357361555, + 0.009627794846892357, + 0.008404800668358803, + 0.006277670618146658, + 0.006474513094872236, + 0.014074261300265789, + 0.012190969660878181, + 0.005881601944565773, + 0.010852303355932236, + 0.0331265814602375, + 0.0959395095705986, + 0.06931571662425995, + 0.014076963998377323, + 0.019731814041733742, + 0.010120626538991928, + 0.0065970090217888355, + 0.011156602762639523, + 0.01463348139077425, + 0.023629536852240562, + 0.012573120184242725, + 0.009893771260976791, + 0.014719591476023197, + 0.02072281949222088, + 0.011945747770369053, + 0.004401191137731075, + 0.07033885270357132, + 0.03752315416932106, + 0.00907664094120264, + 0.019513333216309547, + 0.009155980311334133, + 0.04255754128098488, + 0.003771024290472269, + 0.005654679145663977, + 0.0032868562266230583, + 0.004439988639205694, + 0.0076837074011564255, + 0.008318508975207806, + 0.004889707081019878, + 0.00723770447075367, + 0.014885726384818554, + 0.00994462426751852, + 0.010678218677639961, + 0.011294319294393063, + 0.02863447740674019, + 0.017575562000274658, + 0.010779748670756817, + 0.008996164426207542, + 0.001759195583872497, + 0.001866622013039887, + 0.001845539198257029, + 0.0020285260397940874, + 0.0035394744481891394, + 0.0029955573845654726, + 0.0018819710239768028, + 0.002943921834230423, + 0.0012762256665155292, + 0.0019680969417095184, + 0.0011886386200785637, + 0.0021107150241732597, + 0.0014662222238257527, + 0.0016524064121767879, + 0.001383305643685162, + 0.0015708762221038342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.00865836814045906 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003137469291687 + ], + "opposite_logprob_tokens": [ + -2.6950314044952393 + ], + "logprob_reward": -0.07003137469291687, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003137469291687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13029146194458008, + 0.041673582047224045, + 0.05275040492415428, + 0.10203128308057785, + 0.04528424143791199, + 0.29382526874542236, + 0.09584040939807892, + 0.19042159616947174, + 0.031223297119140625, + 0.016658425331115723 + ], + "credit_map": [ + 0.010114145465195179, + 0.007248304318636656, + 0.02189473807811737, + 0.028678324073553085, + 0.013782002963125706, + 0.008883727714419365, + 0.016813354566693306, + 0.022876858711242676, + 0.0035793764982372522, + 0.002603271510452032, + 0.006606840528547764, + 0.007916565984487534, + 0.004420551471412182, + 0.004127446562051773, + 0.005498442333191633, + 0.006921086926013231, + 0.004021772183477879, + 0.004545002710074186, + 0.009115468710660934, + 0.010288532823324203, + 0.003785221604630351, + 0.005214990116655827, + 0.007583838887512684, + 0.008195582777261734, + 0.0059016309678554535, + 0.004390442743897438, + 0.013364851474761963, + 0.034611281007528305, + 0.021527739241719246, + 0.006968650035560131, + 0.008216227404773235, + 0.00705046346411109, + 0.002627199748530984, + 0.0039009754545986652, + 0.00934654287993908, + 0.008533773012459278, + 0.006140039302408695, + 0.004058687016367912, + 0.00568764004856348, + 0.004989382810890675, + 0.019259795546531677, + 0.01086975447833538, + 0.09845804423093796, + 0.05508920177817345, + 0.016557389870285988, + 0.02551238052546978, + 0.01347682997584343, + 0.05460185930132866, + 0.00646916963160038, + 0.01878521777689457, + 0.005700006615370512, + 0.00900881178677082, + 0.02230183035135269, + 0.01548545528203249, + 0.007268446031957865, + 0.010821476578712463, + 0.023178985342383385, + 0.015558719635009766, + 0.019989095628261566, + 0.01859857514500618, + 0.03995164856314659, + 0.030003702268004417, + 0.022057391703128815, + 0.02108348347246647, + 0.004125313833355904, + 0.003226365428417921, + 0.003064251272007823, + 0.0037715856451541185, + 0.00429965415969491, + 0.004522015806287527, + 0.0031629600562155247, + 0.005051152780652046, + 0.0014355093007907271, + 0.0020551374182105064, + 0.0015918652061372995, + 0.00264334911480546, + 0.0019763486925512552, + 0.0023439566139131784, + 0.0020048704463988543, + 0.002607388189062476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.07003137469291687 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978246983140707 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.00978246983140707, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978246983140707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12856146693229675, + 0.03039824217557907, + 0.023214468732476234, + 0.1471370905637741, + 0.06620030850172043, + 0.11560159921646118, + 0.19562850892543793, + 0.20272105932235718, + 0.05113793537020683, + 0.039399344474077225 + ], + "credit_map": [ + 0.011060885153710842, + 0.006778999697417021, + 0.01197475753724575, + 0.038924504071474075, + 0.015770481899380684, + 0.007845012471079826, + 0.014572037383913994, + 0.021634787321090698, + 0.0034389568027108908, + 0.002173928078263998, + 0.005301777273416519, + 0.004079234320670366, + 0.0024788030423223972, + 0.0030784548725932837, + 0.004503863397985697, + 0.005343222059309483, + 0.002937040291726589, + 0.0024046797771006823, + 0.0026053760666400194, + 0.0022403253242373466, + 0.0023853499442338943, + 0.002848041011020541, + 0.004298466257750988, + 0.0034951891284435987, + 0.004731335677206516, + 0.006915920414030552, + 0.017773017287254333, + 0.0526447594165802, + 0.040334202349185944, + 0.008634917438030243, + 0.010830674320459366, + 0.005272245965898037, + 0.003160415915772319, + 0.00398391904309392, + 0.01976310834288597, + 0.008959200233221054, + 0.011701937764883041, + 0.006873955484479666, + 0.006025377660989761, + 0.00573238730430603, + 0.00542783597484231, + 0.008961010724306107, + 0.028799692168831825, + 0.01448019314557314, + 0.012866757810115814, + 0.01642434485256672, + 0.008053944446146488, + 0.02058781124651432, + 0.014252702705562115, + 0.03924889862537384, + 0.013478401117026806, + 0.020912734791636467, + 0.04116953909397125, + 0.03098570741713047, + 0.01462646946310997, + 0.020954038947820663, + 0.025404388085007668, + 0.018190328031778336, + 0.019325220957398415, + 0.020014263689517975, + 0.050326358526945114, + 0.031011085957288742, + 0.019914796575903893, + 0.01853460632264614, + 0.007718782406300306, + 0.005266958847641945, + 0.005400857888162136, + 0.006162685342133045, + 0.005546014290302992, + 0.006891966797411442, + 0.005351278465241194, + 0.008799385279417038, + 0.003962249495089054, + 0.0058319540694355965, + 0.003704649396240711, + 0.006775230634957552, + 0.005397011525928974, + 0.005032330751419067, + 0.0038879604544490576, + 0.004807954654097557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.00978246983140707 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle focusing on a corner of a dark solid wood desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a printed diagnosis report with glaring white paper lying neatly spread out on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pair of unused reading glasses visible beside the report?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does an elderly hand with wrinkled, lightly freckled skin enter the frame and tremble while touching the paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly move forward and then zoom in to show the details of the finger?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864705726504326 + ], + "opposite_logprob_tokens": [ + -3.048647165298462 + ], + "logprob_reward": -0.04864705726504326, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864705726504326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20655003190040588, + 0.045230135321617126, + 0.056799549609422684, + 0.12390358000993729, + 0.044205572456121445, + 0.15206299722194672, + 0.08995863050222397, + 0.21602055430412292, + 0.030743759125471115, + 0.034525156021118164 + ], + "credit_map": [ + 0.012700078077614307, + 0.012444403022527695, + 0.020313400775194168, + 0.045707959681749344, + 0.020950978621840477, + 0.01790301501750946, + 0.03418849781155586, + 0.04234171286225319, + 0.004720285069197416, + 0.00396721763536334, + 0.006992729846388102, + 0.007624926045536995, + 0.005631474778056145, + 0.004643757827579975, + 0.006157490890473127, + 0.00549225602298975, + 0.006517818663269281, + 0.004777637310326099, + 0.006807467434555292, + 0.009928805753588676, + 0.006074568256735802, + 0.00669066933915019, + 0.009488443844020367, + 0.006514145992696285, + 0.007775005418807268, + 0.006237167399376631, + 0.014469217509031296, + 0.039488535374403, + 0.029742566868662834, + 0.008480328135192394, + 0.009602840058505535, + 0.008107932284474373, + 0.002907725516706705, + 0.003703957423567772, + 0.0053801508620381355, + 0.007959790527820587, + 0.008638613857328892, + 0.006531612016260624, + 0.005338464863598347, + 0.0037452576216310263, + 0.009877334348857403, + 0.01122497208416462, + 0.036678094416856766, + 0.020191866904497147, + 0.021828746423125267, + 0.01647094264626503, + 0.008430720306932926, + 0.027360327541828156, + 0.007295406889170408, + 0.015147190541028976, + 0.006027801427990198, + 0.009350838139653206, + 0.02088083140552044, + 0.013304166495800018, + 0.007982654497027397, + 0.009969746693968773, + 0.0259584728628397, + 0.017119647935032845, + 0.021926824003458023, + 0.020855586975812912, + 0.04704417288303375, + 0.03393218293786049, + 0.025460802018642426, + 0.02372286468744278, + 0.004603049252182245, + 0.0038729330990463495, + 0.002969953930005431, + 0.0034166898112744093, + 0.004249625839293003, + 0.004042217507958412, + 0.002967427484691143, + 0.004621864762157202, + 0.003100139321759343, + 0.005225078202784061, + 0.002925089094787836, + 0.006544836331158876, + 0.005502734798938036, + 0.004227562341839075, + 0.003942732699215412, + 0.0030569862574338913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.04864705726504326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_tensor([6917, 6917])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_tensor([6917, 6917])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4995fa38cc10314749f854ff2adc797a6aa402d4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_tensor([6917, 6917])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic high-angle tracking shot follows a vintage biplane as it descents out of control, its camera path spiraling downward to mirror the aircraft's tragic trajectory. The aircraft's fuselage, constructed of weathered, dark green wood and metal, shows signs of distress; the edges of its canvas-covered wings are frayed, and a thin but persistent plume of oily black smoke billows from its engine cowling. On its upper wing, a vibrant circle of vibrant colors—bold concentric circles of faded red, white, and blue—is clearly visible, a stark symbol of its defiance against chaos. The plane plummets not through a normal sky but through a surreal one, filled with impossibly painterly clouds that appear like thick, impassive brushstrokes. These clouds swirl in dramatic swirls of sunset orange, deep purple, and golden ochre, catching the light of a low, unseen sun. The biplane tumbles helplessly, one wing lower than the other in its spiraling fall, its propellers perfectly still. Light filtered through the dreamlike clouds, casting ever-changing patterns of color on the plane’s damaged fuselage and illuminating the black smoke, creating an atmosphere of stunning, silent, and inevitable doom.", + "seed": 31836, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_6917_seed_31836_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020480135921388865, + "vlm_logprob": -0.020480135921388865, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.020480135921388865, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017414521425962448 + ], + "opposite_logprob_tokens": [ + -4.642414569854736 + ], + "logprob_reward": -0.017414521425962448, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.017414521425962448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011425929144024849, + 0.008553816936910152, + 0.8188313841819763, + 0.010528424754738808, + 0.07575926929712296, + 0.008617188781499863, + 0.017710745334625244, + 0.025122253224253654, + 0.013609584420919418, + 0.009841473773121834 + ], + "credit_map": [ + 0.0006931017851457, + 0.0006534283747896552, + 0.0016276355599984527, + 0.0010535229230299592, + 0.0006911739474162459, + 0.0008203585748560727, + 0.002254148945212364, + 0.0036325575783848763, + 0.0010885450756177306, + 0.0010269791819155216, + 0.0011669095838442445, + 0.0015139892930164933, + 0.0008742862264625728, + 0.001096789026632905, + 0.0007629076717421412, + 0.0010234100045636296, + 0.037562910467386246, + 0.010798332281410694, + 0.05082836374640465, + 0.014262580312788486, + 0.024756914004683495, + 0.34179913997650146, + 0.024081410840153694, + 0.3147416412830353, + 0.001595280016772449, + 0.0009745655697770417, + 0.0018374458886682987, + 0.0014878838555887341, + 0.0010084074456244707, + 0.0011609281646087766, + 0.001028307480737567, + 0.0014356047613546252, + 0.005866931285709143, + 0.004178373143076897, + 0.004048592410981655, + 0.0024423222057521343, + 0.00359094375744462, + 0.02581222541630268, + 0.002787176985293627, + 0.027032695710659027, + 0.0014008210273459554, + 0.0007255541277118027, + 0.0012765377759933472, + 0.0006959398742765188, + 0.0008604569011367857, + 0.0011306938249617815, + 0.0008925746660679579, + 0.0016346090706065297, + 0.0035666711628437042, + 0.001208648318424821, + 0.0025464221835136414, + 0.0019066499080508947, + 0.001318843918852508, + 0.002093259245157242, + 0.00257957074791193, + 0.0024906776379793882, + 0.0020081435795873404, + 0.0010147029533982277, + 0.006988546345382929, + 0.0011144299060106277, + 0.001292928820475936, + 0.009168077260255814, + 0.0017963747959583998, + 0.0017390461871400476, + 0.0024506570771336555, + 0.0014626129996031523, + 0.0016441663028672338, + 0.0019858803134411573, + 0.0011934156063944101, + 0.0019740224815905094, + 0.00129705888684839, + 0.0016017691232264042, + 0.0012501893797889352, + 0.0011477551888674498, + 0.0021254008170217276, + 0.0010637847008183599, + 0.0008811000734567642, + 0.0010587286669760942, + 0.0010885322699323297, + 0.0012259811628609896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.017414521425962448 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370098024606705 + ], + "opposite_logprob_tokens": [ + -3.408700942993164 + ], + "logprob_reward": -0.03370098024606705, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370098024606705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026474131271243095, + 0.003499960293993354, + 0.8858076333999634, + 0.014228780753910542, + 0.020036907866597176, + 0.009901689365506172, + 0.015327407978475094, + 0.008647078648209572, + 0.00659222062677145, + 0.00948425754904747 + ], + "credit_map": [ + 0.001653317711316049, + 0.0018935316475108266, + 0.003348536090925336, + 0.0042437841184437275, + 0.0022882132325321436, + 0.001996453385800123, + 0.004478279035538435, + 0.006572012789547443, + 0.0005345119861885905, + 0.00046138319885358214, + 0.0004386226646602154, + 0.0005274092545732856, + 0.0003505768836475909, + 0.0003746459260582924, + 0.0004120988887734711, + 0.00040071088005788624, + 0.04088844358921051, + 0.011856469325721264, + 0.054752644151449203, + 0.01571609079837799, + 0.026924854144454002, + 0.37007853388786316, + 0.026300596073269844, + 0.33928993344306946, + 0.001958301989361644, + 0.0014131634961813688, + 0.0025685306172817945, + 0.001432624994777143, + 0.0016293817898258567, + 0.0018960749730467796, + 0.0009810030460357666, + 0.002349698217585683, + 0.0016901433700695634, + 0.0011217236751690507, + 0.0013248971663415432, + 0.0007431500707753003, + 0.001166907837614417, + 0.006127459462732077, + 0.0011025475105270743, + 0.006760077551007271, + 0.001416030921973288, + 0.0011378575582057238, + 0.0014051944017410278, + 0.0007864087237976491, + 0.0012429497437551618, + 0.0015396183589473367, + 0.001210537040606141, + 0.0011630906956270337, + 0.002708256943151355, + 0.001253417576663196, + 0.0021537768188863993, + 0.0016743907472118735, + 0.0009670404833741486, + 0.001865209313109517, + 0.0022059776820242405, + 0.0024993373081088066, + 0.0009822663851082325, + 0.0004636724479496479, + 0.00261936173774302, + 0.0004717758856713772, + 0.00045379161019809544, + 0.0025873021222651005, + 0.00048758272896520793, + 0.0005813244497403502, + 0.0008023004047572613, + 0.0005014244816266, + 0.000896332785487175, + 0.0006789996987208724, + 0.000913160911295563, + 0.000906284898519516, + 0.0011315859155729413, + 0.0007621304248459637, + 0.0015582924243062735, + 0.0010181271936744452, + 0.0020933549385517836, + 0.0008404610562138259, + 0.0010375853162258863, + 0.0013196875806897879, + 0.0007671507191844285, + 0.0008495970978401601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.03370098024606705 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579854026436806 + ], + "opposite_logprob_tokens": [ + -3.8955798149108887 + ], + "logprob_reward": -0.020579854026436806, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579854026436806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.007394012063741684, + 0.0044131348840892315, + 0.5619059801101685, + 0.0073595005087554455, + 0.33590391278266907, + 0.006764841265976429, + 0.0048923185095191, + 0.024967651814222336, + 0.03965223208069801, + 0.006746390368789434 + ], + "credit_map": [ + 0.0005115395761094987, + 0.0008707437664270401, + 0.001008410006761551, + 0.0010946313850581646, + 0.0007564220577478409, + 0.0008193803369067609, + 0.0009353875066153705, + 0.0013974980683997273, + 0.0010414995485916734, + 0.0004575340135488659, + 0.0007415569270960987, + 0.00040938309393823147, + 0.0003774371580220759, + 0.0005055864457972348, + 0.0003484877815935761, + 0.000531650148332119, + 0.025750918313860893, + 0.007473276928067207, + 0.03494838997721672, + 0.009991290979087353, + 0.017066629603505135, + 0.23427991569042206, + 0.016669830307364464, + 0.21572577953338623, + 0.0011350596323609352, + 0.0004818942106794566, + 0.001389437005855143, + 0.0008239862509071827, + 0.0006034887628629804, + 0.0009261747472919524, + 0.0006400927668437362, + 0.0013593672774732113, + 0.014611119404435158, + 0.0079470444470644, + 0.018443895503878593, + 0.004841332323849201, + 0.011004174128174782, + 0.1282963901758194, + 0.006060316227376461, + 0.14469967782497406, + 0.0008608796051703393, + 0.0004873317666351795, + 0.0011424614349380136, + 0.0005235928110778332, + 0.0005220382008701563, + 0.001415538601577282, + 0.0007342894678004086, + 0.0010787103092297912, + 0.0008319345070049167, + 0.0007350922678597271, + 0.0005725991213694215, + 0.0004846078809350729, + 0.00033741985680535436, + 0.0006203242228366435, + 0.00046625672257505357, + 0.0008440843084827065, + 0.0019270469201728702, + 0.0006020274013280869, + 0.008070283569395542, + 0.0010834430577233434, + 0.0012485557235777378, + 0.009821051731705666, + 0.0011344141094014049, + 0.0010808334918692708, + 0.0055891177617013454, + 0.004453062079846859, + 0.00482935318723321, + 0.005218130070716143, + 0.004437910858541727, + 0.00475542526692152, + 0.00523448595777154, + 0.005134751554578543, + 0.0009124873322434723, + 0.0008197343559004366, + 0.00106070376932621, + 0.0007295829127542675, + 0.0007471233257092535, + 0.000785370881203562, + 0.0008574661333113909, + 0.000833922007586807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.020579854026436806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029786404222249985 + ], + "opposite_logprob_tokens": [ + -3.5297863483428955 + ], + "logprob_reward": -0.029786404222249985, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029786404222249985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.007415436208248138, + 0.0030945129692554474, + 0.9315090775489807, + 0.0062788124196231365, + 0.024638015776872635, + 0.002889950294047594, + 0.006593836471438408, + 0.0039004390127956867, + 0.007310385815799236, + 0.006369438488036394 + ], + "credit_map": [ + 0.0005828236462548375, + 0.000684463360812515, + 0.0011648080544546247, + 0.001364838914014399, + 0.0008301506168209016, + 0.0006193859153427184, + 0.0013995227636769414, + 0.0007694425294175744, + 0.0005197690334171057, + 0.00032012027804739773, + 0.00047214978258125484, + 0.00045391361345537007, + 0.00023833202430978417, + 0.0003353761276230216, + 0.0002811729209497571, + 0.0004736793343909085, + 0.042905986309051514, + 0.012313606217503548, + 0.05775129050016403, + 0.016574449837207794, + 0.028278687968850136, + 0.3887903392314911, + 0.027569059282541275, + 0.3573257327079773, + 0.0009032862726598978, + 0.00047371149412356317, + 0.0011713297571986914, + 0.0008510389598086476, + 0.0004950030124746263, + 0.0007310976507142186, + 0.0006396639510057867, + 0.00101368164177984, + 0.0018060366855934262, + 0.000725698540918529, + 0.0013250214979052544, + 0.0005671594408340752, + 0.0010171537287533283, + 0.009056716226041317, + 0.0006207434926182032, + 0.009519485756754875, + 0.00042037595994770527, + 0.00026141523267142475, + 0.0004014632140751928, + 0.000293461955152452, + 0.0002717288734856993, + 0.0003607877588365227, + 0.0005033236229792237, + 0.00037739373510703444, + 0.0007350604864768684, + 0.0003392898361198604, + 0.0015317968791350722, + 0.0004551617312245071, + 0.0003552960406523198, + 0.001396005623973906, + 0.0005830902373418212, + 0.0011981356656178832, + 0.00043735888903029263, + 0.0003008166095241904, + 0.0009647923870943487, + 0.00023873943428043276, + 0.00019550799333956093, + 0.0011640406446531415, + 0.00029301701579242945, + 0.00030616624280810356, + 0.0009689321159385145, + 0.0011104629375040531, + 0.0008487256127409637, + 0.0007912422297522426, + 0.0009641546057537198, + 0.0009368755272589624, + 0.0006681952509097755, + 0.001021797303110361, + 0.0006851968355476856, + 0.0006766124861314893, + 0.0009928005747497082, + 0.0008381091756746173, + 0.0008127276669256389, + 0.0006313459598459303, + 0.0009157339809462428, + 0.0008169112261384726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.029786404222249985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009189196862280369 + ], + "opposite_logprob_tokens": [ + -7.000918865203857 + ], + "logprob_reward": -0.0009189196862280369, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009189196862280369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016017552465200424, + 0.0068414597772061825, + 0.8643438220024109, + 0.009608681313693523, + 0.05852540582418442, + 0.008279206231236458, + 0.009911862201988697, + 0.014300269074738026, + 0.007382858078926802, + 0.004788899794220924 + ], + "credit_map": [ + 0.002307636896148324, + 0.0017426322447136045, + 0.002304145833477378, + 0.0032274278346449137, + 0.0014300851617008448, + 0.0018201322527602315, + 0.0013248580507934093, + 0.001860633259639144, + 0.0012864761520177126, + 0.0010239190887659788, + 0.0009465821203775704, + 0.0008759051561355591, + 0.0004535975749604404, + 0.0008582731825299561, + 0.0005434544291347265, + 0.0008532522479072213, + 0.0395951084792614, + 0.011520208790898323, + 0.05376961827278137, + 0.01541179046034813, + 0.02618705853819847, + 0.3604581356048584, + 0.025625906884670258, + 0.3317760229110718, + 0.0009298050426878035, + 0.001026310957968235, + 0.0015863848384469748, + 0.001178357983008027, + 0.0011892786715179682, + 0.001256366609595716, + 0.0008046855218708515, + 0.0016374916303902864, + 0.003834115108475089, + 0.0022586400154978037, + 0.0045160274021327496, + 0.002253544284030795, + 0.002402746817097068, + 0.019652383401989937, + 0.0027252761647105217, + 0.02088267356157303, + 0.0011398948263376951, + 0.0009550079121254385, + 0.001260802848264575, + 0.0006216246983967721, + 0.0009654993773438036, + 0.0013653297210112214, + 0.0009118410525843501, + 0.0010592059697955847, + 0.0010228105820715427, + 0.0006070335512049496, + 0.0021555901039391756, + 0.000920026155654341, + 0.0005895013455301523, + 0.0014044448034837842, + 0.0013575175544247031, + 0.0018549386877566576, + 0.0013358264695852995, + 0.0005482507403939962, + 0.004206574056297541, + 0.0006193271256051958, + 0.0007237847312353551, + 0.005226494744420052, + 0.0007864679791964591, + 0.000853543751873076, + 0.0012510234955698252, + 0.0008233899716287851, + 0.0009680513176135719, + 0.0007047862163744867, + 0.0008530649938620627, + 0.0009306491119787097, + 0.0009365053265355527, + 0.0009153874707408249, + 0.0007821032777428627, + 0.0006018200074322522, + 0.000768228666856885, + 0.0005742661887779832, + 0.0006455074762925506, + 0.0004941922379657626, + 0.00042714737355709076, + 0.0004956342163495719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.0009189196862280369 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic high-angle tracking shot follows a vintage biplane as it descents out of control, its camera path spiraling downward to mirror the aircraft's tragic trajectory. The aircraft's fuselage, constructed of weathered, dark green wood and metal, shows signs of distress; the edges of its canvas-covered wings are frayed, and a thin but persistent plume of oily black smoke billows from its engine cowling. On its upper wing, a vibrant circle of vibrant colors—bold concentric circles of faded red, white, and blue—is clearly visible, a stark symbol of its defiance against chaos. The plane plummets not through a normal sky but through a surreal one, filled with impossibly painterly clouds that appear like thick, impassive brushstrokes. These clouds swirl in dramatic swirls of sunset orange, deep purple, and golden ochre, catching the light of a low, unseen sun. The biplane tumbles helplessly, one wing lower than the other in its spiraling fall, its propellers perfectly still. Light filtered through the dreamlike clouds, casting ever-changing patterns of color on the plane’s damaged fuselage and illuminating the black smoke, creating an atmosphere of stunning, silent, and inevitable doom.", + "seed": 31837, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_6917_seed_31837_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020107717579230668, + "vlm_logprob": -0.020107717579230668, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.020107717579230668, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014436440542340279 + ], + "opposite_logprob_tokens": [ + -4.6394362449646 + ], + "logprob_reward": -0.014436440542340279, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.014436440542340279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13101617991924286, + 0.05422702431678772, + 0.2442430853843689, + 0.040373146533966064, + 0.03645904362201691, + 0.04850568249821663, + 0.13354556262493134, + 0.23788517713546753, + 0.03541292995214462, + 0.03833220526576042 + ], + "credit_map": [ + 0.012806831859052181, + 0.012021037749946117, + 0.019060246646404266, + 0.017194386571645737, + 0.012088490650057793, + 0.022700460627675056, + 0.013614051975309849, + 0.02153065986931324, + 0.004400245845317841, + 0.0043196603655815125, + 0.011545869521796703, + 0.005198122933506966, + 0.004579448141157627, + 0.009815040975809097, + 0.0034047465305775404, + 0.010963885113596916, + 0.012788630090653896, + 0.01796908676624298, + 0.022531896829605103, + 0.015372902154922485, + 0.07181250303983688, + 0.029797296971082687, + 0.011872005648911, + 0.06209873408079147, + 0.005421695299446583, + 0.004226970952004194, + 0.004366024862974882, + 0.0037057872395962477, + 0.005056638736277819, + 0.006733366753906012, + 0.005747826304286718, + 0.005114835686981678, + 0.004514518193900585, + 0.0035767259541898966, + 0.002212564228102565, + 0.002713762456551194, + 0.005402205046266317, + 0.005667057819664478, + 0.0038698967546224594, + 0.008502311073243618, + 0.007385472767055035, + 0.004735651891678572, + 0.004499016795307398, + 0.0045822965912520885, + 0.005193884018808603, + 0.006328665651381016, + 0.0050599705427885056, + 0.01072071772068739, + 0.013891327194869518, + 0.042536765336990356, + 0.00573225412517786, + 0.005442965310066938, + 0.004707388114184141, + 0.041156306862831116, + 0.003041223855689168, + 0.017037319019436836, + 0.03069767914712429, + 0.02668047696352005, + 0.08419676125049591, + 0.006007389631122351, + 0.006204071454703808, + 0.021758602932095528, + 0.00509134354069829, + 0.057248834520578384, + 0.004794285632669926, + 0.0051221661269664764, + 0.004466252401471138, + 0.005171176977455616, + 0.0038620734121650457, + 0.0036310271825641394, + 0.004535131622105837, + 0.003830812405794859, + 0.005364954471588135, + 0.0036451451014727354, + 0.004238104913383722, + 0.004220296163111925, + 0.004630293231457472, + 0.005294534843415022, + 0.0049360101111233234, + 0.006002862472087145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -0.014436440542340279 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785421192646027 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.009785421192646027, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785421192646027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07444954663515091, + 0.023355955258011818, + 0.09354645013809204, + 0.07081791758537292, + 0.060615308582782745, + 0.025225643068552017, + 0.061043668538331985, + 0.49107033014297485, + 0.05925550311803818, + 0.04061964899301529 + ], + "credit_map": [ + 0.008674612268805504, + 0.0052437311969697475, + 0.011902142316102982, + 0.008396865800023079, + 0.005016767419874668, + 0.016222970560193062, + 0.006846034433692694, + 0.012146428227424622, + 0.002213728614151478, + 0.0016516383038833737, + 0.00605433713644743, + 0.002671366324648261, + 0.0023446278646588326, + 0.003704310627654195, + 0.0011496972292661667, + 0.0035662518348544836, + 0.007983195595443249, + 0.011679085902869701, + 0.00967214535921812, + 0.01573188230395317, + 0.011547130532562733, + 0.015340413898229599, + 0.01107058860361576, + 0.010522016324102879, + 0.0070596421137452126, + 0.008348190225660801, + 0.00791194848716259, + 0.00685886712744832, + 0.01020064391195774, + 0.009905554354190826, + 0.010270597413182259, + 0.010262482799589634, + 0.007799390237778425, + 0.007205488160252571, + 0.006050299387425184, + 0.005862618330866098, + 0.0110006183385849, + 0.0074488078244030476, + 0.006919241975992918, + 0.00832885131239891, + 0.003263516817241907, + 0.004150370601564646, + 0.00195444212295115, + 0.002724020043388009, + 0.002332363510504365, + 0.003605585778132081, + 0.0021122111938893795, + 0.005083134397864342, + 0.010389307513833046, + 0.015223334543406963, + 0.0039303298108279705, + 0.002889741212129593, + 0.0035142030101269484, + 0.010347058065235615, + 0.002849743701517582, + 0.011899951845407486, + 0.06359737366437912, + 0.055314768105745316, + 0.17670926451683044, + 0.01186456810683012, + 0.011633663438260555, + 0.043072979897260666, + 0.010600338689982891, + 0.11827739328145981, + 0.004948439076542854, + 0.00525007164105773, + 0.009082783944904804, + 0.011193160898983479, + 0.0063670785166323185, + 0.00777492206543684, + 0.007867861539125443, + 0.006771191023290157, + 0.007456357125192881, + 0.004177016206085682, + 0.004837933927774429, + 0.006266594864428043, + 0.003744719084352255, + 0.004315228667110205, + 0.004510792437940836, + 0.0053110080771148205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.009785421192646027 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549495592713356 + ], + "opposite_logprob_tokens": [ + -2.9299495220184326 + ], + "logprob_reward": -0.0549495592713356, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549495592713356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0864575132727623, + 0.05988610163331032, + 0.296628475189209, + 0.060912538319826126, + 0.11349799484014511, + 0.107438825070858, + 0.015398687683045864, + 0.13482968509197235, + 0.09127368777990341, + 0.03367649391293526 + ], + "credit_map": [ + 0.007325875572860241, + 0.009278758428990841, + 0.010262877680361271, + 0.011278921738266945, + 0.0061654141172766685, + 0.015221727080643177, + 0.006270056124776602, + 0.020653875544667244, + 0.0050072637386620045, + 0.00451089208945632, + 0.015301566570997238, + 0.005072431173175573, + 0.005026103928685188, + 0.00864199735224247, + 0.004815413150936365, + 0.011510435491800308, + 0.01017184928059578, + 0.010271856561303139, + 0.01655551791191101, + 0.01143084466457367, + 0.10466813296079636, + 0.03530517593026161, + 0.01021370105445385, + 0.09801140427589417, + 0.005867627914994955, + 0.005956088192760944, + 0.00625073816627264, + 0.006119271274656057, + 0.00983030442148447, + 0.010032967664301395, + 0.0079775620251894, + 0.008877977728843689, + 0.01937894895672798, + 0.009567514061927795, + 0.009814463555812836, + 0.006494090426713228, + 0.010970941744744778, + 0.018778640776872635, + 0.01697145402431488, + 0.02152193896472454, + 0.01763271354138851, + 0.009281295351684093, + 0.007839545607566833, + 0.006374277640134096, + 0.009422024711966515, + 0.01581779308617115, + 0.01257151085883379, + 0.02849966660141945, + 0.0023794302251189947, + 0.002057911828160286, + 0.0013059080811217427, + 0.0010661846026778221, + 0.0009311707108281553, + 0.002600656356662512, + 0.0007276851101778448, + 0.00432974100112915, + 0.017807792872190475, + 0.016555098816752434, + 0.04695368930697441, + 0.003658213187009096, + 0.0035236640833318233, + 0.012245621532201767, + 0.0033300810027867556, + 0.03075552172958851, + 0.013667489401996136, + 0.009583022445440292, + 0.011202286928892136, + 0.018593566492199898, + 0.01067415252327919, + 0.009808971546590328, + 0.008634166792035103, + 0.009110032580792904, + 0.005507203750312328, + 0.0034474374260753393, + 0.004745288752019405, + 0.004003526642918587, + 0.002731352113187313, + 0.005286330822855234, + 0.0033187714871019125, + 0.004636585712432861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.0549495592713356 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178876489400864 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178876489400864, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178876489400864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07438638806343079, + 0.05107938125729561, + 0.07502108067274094, + 0.05939541384577751, + 0.11145398765802383, + 0.05426209047436714, + 0.1888366937637329, + 0.2378375232219696, + 0.07053191214799881, + 0.07719559967517853 + ], + "credit_map": [ + 0.009862616658210754, + 0.004617036320269108, + 0.012238473631441593, + 0.011756166815757751, + 0.004890069831162691, + 0.012387208640575409, + 0.006970503833144903, + 0.011664312332868576, + 0.004585085436701775, + 0.005314054898917675, + 0.00845203548669815, + 0.006274050567299128, + 0.004560309462249279, + 0.008444597944617271, + 0.0036095886025577784, + 0.009839660488069057, + 0.005483099725097418, + 0.005722660105675459, + 0.006702054757624865, + 0.0043943640775978565, + 0.019534114748239517, + 0.011922112666070461, + 0.004852913320064545, + 0.01640976034104824, + 0.004480844829231501, + 0.005184653215110302, + 0.0045161074958741665, + 0.005027638282626867, + 0.008299808017909527, + 0.010640175081789494, + 0.010487910360097885, + 0.010758274234831333, + 0.013918041251599789, + 0.015483054332435131, + 0.012444322928786278, + 0.01506058033555746, + 0.01676071248948574, + 0.014091944321990013, + 0.011590912006795406, + 0.012104419991374016, + 0.005541904363781214, + 0.005794795695692301, + 0.006276407744735479, + 0.006404302082955837, + 0.007243995554745197, + 0.0068231020122766495, + 0.007569603621959686, + 0.00860797893255949, + 0.010009421035647392, + 0.0541069470345974, + 0.006176406983286142, + 0.007316359784454107, + 0.00664213253185153, + 0.07463949173688889, + 0.0029163872823119164, + 0.027029527351260185, + 0.0297856442630291, + 0.029008179903030396, + 0.08497760444879532, + 0.005391784943640232, + 0.004968797788023949, + 0.021252349019050598, + 0.0049851094372570515, + 0.05746803060173988, + 0.006451021879911423, + 0.0068111298605799675, + 0.006838929373770952, + 0.016408458352088928, + 0.008864223957061768, + 0.007798471488058567, + 0.010199581272900105, + 0.007160099223256111, + 0.011407085694372654, + 0.007998107001185417, + 0.010009020566940308, + 0.010814618319272995, + 0.009319134056568146, + 0.011160477064549923, + 0.008861737325787544, + 0.0076254140585660934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.018178876489400864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00318829040043056 + ], + "opposite_logprob_tokens": [ + -5.753188133239746 + ], + "logprob_reward": -0.00318829040043056, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00318829040043056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037619441747665405, + 0.0290264543145895, + 0.24578620493412018, + 0.10215064883232117, + 0.04732879251241684, + 0.0733538568019867, + 0.08282829821109772, + 0.26970136165618896, + 0.06713782250881195, + 0.04506710544228554 + ], + "credit_map": [ + 0.003061319701373577, + 0.0037748413160443306, + 0.004828234203159809, + 0.0038646969478577375, + 0.0030838565435260534, + 0.008197452872991562, + 0.0027332045137882233, + 0.008075839839875698, + 0.002290862612426281, + 0.0026982417330145836, + 0.007042294833809137, + 0.0034421896561980247, + 0.0026470343582332134, + 0.0043244073167443275, + 0.002649021567776799, + 0.003932406194508076, + 0.008337194100022316, + 0.006467597559094429, + 0.016279764473438263, + 0.01162186823785305, + 0.0792524442076683, + 0.03794293478131294, + 0.010581930167973042, + 0.07530248910188675, + 0.010480018332600594, + 0.009958090260624886, + 0.012457441538572311, + 0.011389194056391716, + 0.01497477013617754, + 0.012353483587503433, + 0.01888018101453781, + 0.011657476425170898, + 0.00647246278822422, + 0.004273285623639822, + 0.004425798077136278, + 0.003710507182404399, + 0.006151974201202393, + 0.007090283557772636, + 0.0073380907997488976, + 0.007866394706070423, + 0.011940672993659973, + 0.005831551738083363, + 0.0040124994702637196, + 0.00375863304361701, + 0.00554305175319314, + 0.011393703520298004, + 0.007873977534472942, + 0.022999776527285576, + 0.010066542774438858, + 0.02425142377614975, + 0.0029124338179826736, + 0.0032631875947117805, + 0.0032788333483040333, + 0.025889934971928596, + 0.0019381317542865872, + 0.011227820999920368, + 0.04006844013929367, + 0.03333142772316933, + 0.07638518512248993, + 0.015416456386446953, + 0.012066739611327648, + 0.035820122808218, + 0.012078814208507538, + 0.04453418776392937, + 0.011938984505832195, + 0.009451229125261307, + 0.008119878359138966, + 0.00761650875210762, + 0.005170735064893961, + 0.006920583546161652, + 0.007061859127134085, + 0.010858045890927315, + 0.007501669693738222, + 0.003827271517366171, + 0.005051483865827322, + 0.003760828170925379, + 0.0042238556779921055, + 0.006549193523824215, + 0.0047447714023292065, + 0.00940803624689579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.00318829040043056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic high-angle tracking shot follows a vintage biplane as it descents out of control, its camera path spiraling downward to mirror the aircraft's tragic trajectory. The aircraft's fuselage, constructed of weathered, dark green wood and metal, shows signs of distress; the edges of its canvas-covered wings are frayed, and a thin but persistent plume of oily black smoke billows from its engine cowling. On its upper wing, a vibrant circle of vibrant colors—bold concentric circles of faded red, white, and blue—is clearly visible, a stark symbol of its defiance against chaos. The plane plummets not through a normal sky but through a surreal one, filled with impossibly painterly clouds that appear like thick, impassive brushstrokes. These clouds swirl in dramatic swirls of sunset orange, deep purple, and golden ochre, catching the light of a low, unseen sun. The biplane tumbles helplessly, one wing lower than the other in its spiraling fall, its propellers perfectly still. Light filtered through the dreamlike clouds, casting ever-changing patterns of color on the plane’s damaged fuselage and illuminating the black smoke, creating an atmosphere of stunning, silent, and inevitable doom.", + "seed": 31838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_6917_seed_31838_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017568781087175012, + "vlm_logprob": -0.017568781087175012, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.017568781087175012, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012888934463262558 + ], + "opposite_logprob_tokens": [ + -4.8878889083862305 + ], + "logprob_reward": -0.012888934463262558, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012888934463262558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2076806128025055, + 0.07821023464202881, + 0.043355029076337814, + 0.06250912696123123, + 0.08165082335472107, + 0.2143271416425705, + 0.11238858103752136, + 0.04378243163228035, + 0.0643899142742157, + 0.09170609712600708 + ], + "credit_map": [ + 0.03375472128391266, + 0.04045936092734337, + 0.03844934329390526, + 0.02138856053352356, + 0.01887170784175396, + 0.019681939855217934, + 0.01848704367876053, + 0.01658795028924942, + 0.010310082696378231, + 0.011977522633969784, + 0.009746626019477844, + 0.009274782612919807, + 0.009324425831437111, + 0.006950948387384415, + 0.00959048978984356, + 0.011035355739295483, + 0.005714691709727049, + 0.005208109505474567, + 0.007572629489004612, + 0.005992499645799398, + 0.005091219674795866, + 0.006337456870824099, + 0.0032370341941714287, + 0.004201384726911783, + 0.01014585793018341, + 0.008985377848148346, + 0.008875753730535507, + 0.01065773144364357, + 0.005611348431557417, + 0.004863198846578598, + 0.004216526634991169, + 0.009153332561254501, + 0.015212028287351131, + 0.010336247272789478, + 0.01203970704227686, + 0.010168100707232952, + 0.007410927675664425, + 0.011977991089224815, + 0.006410330068320036, + 0.008095492608845234, + 0.0644025057554245, + 0.034106895327568054, + 0.03728969767689705, + 0.017288243398070335, + 0.013997592031955719, + 0.011582709848880768, + 0.012281764298677444, + 0.02337775193154812, + 0.01755797304213047, + 0.011430812999606133, + 0.017598161473870277, + 0.014409845694899559, + 0.011321209371089935, + 0.012406299822032452, + 0.010097768157720566, + 0.017566511407494545, + 0.005056951195001602, + 0.006084217224270105, + 0.005064374767243862, + 0.007461147848516703, + 0.003472310723736882, + 0.005887381266802549, + 0.003910197876393795, + 0.006845853291451931, + 0.009902513585984707, + 0.010340007953345776, + 0.008228261023759842, + 0.00774784991517663, + 0.006602242588996887, + 0.007312360219657421, + 0.006617957726120949, + 0.007638720795512199, + 0.013153154402971268, + 0.011019843630492687, + 0.017392583191394806, + 0.00808023288846016, + 0.00842637661844492, + 0.01051242370158434, + 0.008005350828170776, + 0.015116128139197826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.012888934463262558 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591650158166885 + ], + "opposite_logprob_tokens": [ + -3.8955917358398438 + ], + "logprob_reward": -0.020591650158166885, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591650158166885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11588907986879349, + 0.08376121520996094, + 0.09581410884857178, + 0.032740361988544464, + 0.04337761178612709, + 0.2345115691423416, + 0.19540424644947052, + 0.04728502407670021, + 0.039620183408260345, + 0.11159659922122955 + ], + "credit_map": [ + 0.010503347031772137, + 0.015601133927702904, + 0.01166614517569542, + 0.01627976819872856, + 0.013180781155824661, + 0.01800033077597618, + 0.015095540322363377, + 0.015562032349407673, + 0.012619741261005402, + 0.011537537910044193, + 0.010630212724208832, + 0.01299197319895029, + 0.008656015619635582, + 0.008142289705574512, + 0.007790392264723778, + 0.011393049731850624, + 0.011752759106457233, + 0.00910557433962822, + 0.022546881809830666, + 0.012309592217206955, + 0.007627955172210932, + 0.01788824051618576, + 0.006163568701595068, + 0.008419540710747242, + 0.005455498583614826, + 0.0037527666427195072, + 0.0035752258263528347, + 0.004584367852658033, + 0.0033473738003522158, + 0.003174896351993084, + 0.0033497163094580173, + 0.005500516854226589, + 0.006782081909477711, + 0.007015472277998924, + 0.0099998963996768, + 0.005633956287056208, + 0.0029494394548237324, + 0.0033696398604661226, + 0.003961046226322651, + 0.0036660803016275167, + 0.06611325591802597, + 0.03590761870145798, + 0.03780651465058327, + 0.02325921691954136, + 0.01656498946249485, + 0.012019584886729717, + 0.013193536549806595, + 0.029646839946508408, + 0.059931881725788116, + 0.017703890800476074, + 0.04874298721551895, + 0.01105215959250927, + 0.008918672800064087, + 0.016108691692352295, + 0.006807597819715738, + 0.026138368993997574, + 0.005801700055599213, + 0.005099989008158445, + 0.006642401218414307, + 0.008680704981088638, + 0.00589994341135025, + 0.004338226746767759, + 0.004590108059346676, + 0.006231954786926508, + 0.006732617970556021, + 0.00417547021061182, + 0.0035830631386488676, + 0.004041258245706558, + 0.004753839690238237, + 0.0050781951285898685, + 0.0051057664677500725, + 0.006149974185973406, + 0.016374735161662102, + 0.012986677698791027, + 0.01846042461693287, + 0.009716615080833435, + 0.014263791963458061, + 0.011723831295967102, + 0.010323491878807545, + 0.01774703338742256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.020591650158166885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369844704866409 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.03369844704866409, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369844704866409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042070623487234116, + 0.04689966142177582, + 0.02178662270307541, + 0.02199239656329155, + 0.027196459472179413, + 0.6695692539215088, + 0.06341999769210815, + 0.07176325470209122, + 0.01669257879257202, + 0.018609197810292244 + ], + "credit_map": [ + 0.0069015370681881905, + 0.004252984654158354, + 0.008955078199505806, + 0.003097778419032693, + 0.004208697471767664, + 0.005862576421350241, + 0.004570837132632732, + 0.004221133887767792, + 0.005363957490772009, + 0.008079400286078453, + 0.004385721404105425, + 0.005371614824980497, + 0.007936349138617516, + 0.004853792022913694, + 0.0055996933951973915, + 0.005309132393449545, + 0.0032780165784060955, + 0.0014137641992419958, + 0.005119954701513052, + 0.0019656759686768055, + 0.003033589804545045, + 0.003911460749804974, + 0.0012182623613625765, + 0.0018458990380167961, + 0.0037538395263254642, + 0.002226907527074218, + 0.0033484851010143757, + 0.002950914902612567, + 0.0032188119366765022, + 0.0018339257221668959, + 0.001804123166948557, + 0.0028553884476423264, + 0.005343613680452108, + 0.002937034936621785, + 0.006911527365446091, + 0.0035078965593129396, + 0.002039009239524603, + 0.002073456533253193, + 0.002133091911673546, + 0.002250829478725791, + 0.21256685256958008, + 0.11040903627872467, + 0.11731428653001785, + 0.05214492604136467, + 0.032787587493658066, + 0.03669653832912445, + 0.03266251087188721, + 0.07498747855424881, + 0.015996264293789864, + 0.006975947879254818, + 0.01556483656167984, + 0.004655499942600727, + 0.0033318339847028255, + 0.007025598082691431, + 0.0026516031939536333, + 0.007218416780233383, + 0.009112409316003323, + 0.0073818606324493885, + 0.00874454714357853, + 0.014348733238875866, + 0.006200734060257673, + 0.00851017888635397, + 0.006346109323203564, + 0.011118677444756031, + 0.0023453396279364824, + 0.002722980221733451, + 0.0017203158931806684, + 0.001756673795171082, + 0.0017851389711722732, + 0.0020232440438121557, + 0.0017139537958428264, + 0.0026249324437230825, + 0.0021177036687731743, + 0.00280842836946249, + 0.003518348326906562, + 0.001583665027283132, + 0.002271240344271064, + 0.001970354001969099, + 0.0022726149763911963, + 0.002066842745989561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.03369844704866409 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178876489400864 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178876489400864, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178876489400864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2510126829147339, + 0.055063728243112564, + 0.08017421513795853, + 0.049679774791002274, + 0.04518095403909683, + 0.09456605464220047, + 0.12157559394836426, + 0.11994472146034241, + 0.03906757012009621, + 0.14373467862606049 + ], + "credit_map": [ + 0.02631189487874508, + 0.03521578758955002, + 0.043881140649318695, + 0.029026251286268234, + 0.025899508967995644, + 0.03941316902637482, + 0.026918495073914528, + 0.024346478283405304, + 0.007150697987526655, + 0.0069472393952310085, + 0.0050713615491986275, + 0.005108166486024857, + 0.008117190562188625, + 0.005866545718163252, + 0.006385767366737127, + 0.010416767559945583, + 0.010005740448832512, + 0.010227113030850887, + 0.016017260029911995, + 0.01105405017733574, + 0.007518152706325054, + 0.012770316563546658, + 0.006457553245127201, + 0.006124040111899376, + 0.009173345752060413, + 0.006221405696123838, + 0.007546961307525635, + 0.009889747016131878, + 0.0032424565870314837, + 0.00420322036370635, + 0.0034550230484455824, + 0.005947621539235115, + 0.00759553536772728, + 0.005534911993891001, + 0.009424486197531223, + 0.005074989981949329, + 0.004910345654934645, + 0.005016600713133812, + 0.004201068542897701, + 0.0034230200108140707, + 0.027084149420261383, + 0.016582829877734184, + 0.014542188495397568, + 0.008154308423399925, + 0.007823968306183815, + 0.00514500867575407, + 0.0058940090239048, + 0.009339602664113045, + 0.023295806720852852, + 0.013258918188512325, + 0.041180312633514404, + 0.008120455779135227, + 0.005280185956507921, + 0.009771283715963364, + 0.003939328249543905, + 0.016729312017560005, + 0.013059868477284908, + 0.011023585684597492, + 0.014425709843635559, + 0.016680138185620308, + 0.01627674512565136, + 0.016660993918776512, + 0.013944198377430439, + 0.017873486503958702, + 0.005710645113140345, + 0.004783746786415577, + 0.0034480011090636253, + 0.00338548980653286, + 0.004149303305894136, + 0.005527379456907511, + 0.005166380200535059, + 0.006896626204252243, + 0.013125821016728878, + 0.021090172231197357, + 0.022477027028799057, + 0.012100565247237682, + 0.02422604337334633, + 0.0179708544164896, + 0.016279174014925957, + 0.01646503433585167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.018178876489400864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002485997276380658 + ], + "opposite_logprob_tokens": [ + -6.002486228942871 + ], + "logprob_reward": -0.002485997276380658, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002485997276380658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1441858410835266, + 0.08724410086870193, + 0.011577034369111061, + 0.030896978452801704, + 0.036135368049144745, + 0.20890264213085175, + 0.12236332893371582, + 0.08947744220495224, + 0.12035074830055237, + 0.14886653423309326 + ], + "credit_map": [ + 0.02007465995848179, + 0.011435729451477528, + 0.032115399837493896, + 0.009507241658866405, + 0.010900000110268593, + 0.03112933039665222, + 0.009973381645977497, + 0.019050078466534615, + 0.012382501736283302, + 0.01016051135957241, + 0.014933805912733078, + 0.010237867012619972, + 0.008718881756067276, + 0.008610629476606846, + 0.008618787862360477, + 0.0135811073705554, + 0.0016794156981632113, + 0.0012934742262586951, + 0.001662417664192617, + 0.0014087578747421503, + 0.0017767034005373716, + 0.0012882932787761092, + 0.0010697334073483944, + 0.0013982378877699375, + 0.005823479034006596, + 0.003366083838045597, + 0.004548563156276941, + 0.006215209607034922, + 0.0017554827500134706, + 0.0026608621701598167, + 0.0018151105614379048, + 0.004712186753749847, + 0.006734167691320181, + 0.005091793369501829, + 0.009879565797746181, + 0.004553731065243483, + 0.002331516006961465, + 0.0025436540599912405, + 0.002521771937608719, + 0.002479164395481348, + 0.06268814206123352, + 0.03323440998792648, + 0.03496239706873894, + 0.017813479527831078, + 0.015127332881093025, + 0.010614929720759392, + 0.01106980163604021, + 0.023392153903841972, + 0.023155100643634796, + 0.011094942688941956, + 0.012106525711715221, + 0.006673157215118408, + 0.0062625291757285595, + 0.024425823241472244, + 0.004369121044874191, + 0.03427611663937569, + 0.01327264029532671, + 0.008562515489757061, + 0.013056350871920586, + 0.015118326991796494, + 0.007121472153812647, + 0.011743911541998386, + 0.005684677977114916, + 0.014917544089257717, + 0.02204788103699684, + 0.017972322180867195, + 0.01094553992152214, + 0.01290151383727789, + 0.013301892206072807, + 0.01501227542757988, + 0.011418106965720654, + 0.016751207411289215, + 0.025513559579849243, + 0.014765695668756962, + 0.030006064102053642, + 0.014882584102451801, + 0.01393347978591919, + 0.016743484884500504, + 0.010778644122183323, + 0.022243015468120575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.002485997276380658 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic high-angle tracking shot follows a vintage biplane as it descents out of control, its camera path spiraling downward to mirror the aircraft's tragic trajectory. The aircraft's fuselage, constructed of weathered, dark green wood and metal, shows signs of distress; the edges of its canvas-covered wings are frayed, and a thin but persistent plume of oily black smoke billows from its engine cowling. On its upper wing, a vibrant circle of vibrant colors—bold concentric circles of faded red, white, and blue—is clearly visible, a stark symbol of its defiance against chaos. The plane plummets not through a normal sky but through a surreal one, filled with impossibly painterly clouds that appear like thick, impassive brushstrokes. These clouds swirl in dramatic swirls of sunset orange, deep purple, and golden ochre, catching the light of a low, unseen sun. The biplane tumbles helplessly, one wing lower than the other in its spiraling fall, its propellers perfectly still. Light filtered through the dreamlike clouds, casting ever-changing patterns of color on the plane’s damaged fuselage and illuminating the black smoke, creating an atmosphere of stunning, silent, and inevitable doom.", + "seed": 31839, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_7/index_6917_seed_31839_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07724836679408327, + "vlm_logprob": -0.07724836679408327, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.07724836679408327, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02228107489645481 + ], + "opposite_logprob_tokens": [ + -4.022281169891357 + ], + "logprob_reward": -0.02228107489645481, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02228107489645481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059649139642715454, + 0.03703905642032623, + 0.09017591178417206, + 0.04030013084411621, + 0.04442620649933815, + 0.07740756869316101, + 0.19239065051078796, + 0.3459736108779907, + 0.09109911322593689, + 0.021538572385907173 + ], + "credit_map": [ + 0.013918239623308182, + 0.004343543201684952, + 0.01043798215687275, + 0.01101594790816307, + 0.004291149787604809, + 0.005542994476854801, + 0.0045098657719790936, + 0.0055894204415380955, + 0.005214824341237545, + 0.0031556873582303524, + 0.004750557243824005, + 0.004932088777422905, + 0.007600819692015648, + 0.003967872820794582, + 0.003386003663763404, + 0.0040312036871910095, + 0.010950746014714241, + 0.01223207637667656, + 0.010992872528731823, + 0.00646206084638834, + 0.01897331140935421, + 0.01180445309728384, + 0.00991696398705244, + 0.008843432180583477, + 0.006378109101206064, + 0.004252777434885502, + 0.004921268206089735, + 0.0034933339338749647, + 0.0072899190708994865, + 0.005529079586267471, + 0.0035538391675800085, + 0.004881810396909714, + 0.010069713927805424, + 0.0031810528598725796, + 0.0065643214620649815, + 0.0026893937028944492, + 0.003199071856215596, + 0.006206197664141655, + 0.0031885814387351274, + 0.009327879175543785, + 0.013381273485720158, + 0.006050397176295519, + 0.010571546852588654, + 0.006757686845958233, + 0.010752701200544834, + 0.015068287961184978, + 0.007367940619587898, + 0.007457740604877472, + 0.06540705263614655, + 0.007306505460292101, + 0.08394848555326462, + 0.010073473677039146, + 0.009830769151449203, + 0.004701719619333744, + 0.004557408858090639, + 0.0065652476623654366, + 0.06004886329174042, + 0.023004600778222084, + 0.08041142672300339, + 0.032349325716495514, + 0.03742144629359245, + 0.04927998408675194, + 0.021695489063858986, + 0.04176252335309982, + 0.01075303740799427, + 0.010584061034023762, + 0.011769372969865799, + 0.009448226541280746, + 0.016373226419091225, + 0.010601235553622246, + 0.007336187642067671, + 0.014233770780265331, + 0.002131835324689746, + 0.0021062162704765797, + 0.0017232552636414766, + 0.002253519371151924, + 0.004744984675198793, + 0.002676118863746524, + 0.0020223420578986406, + 0.0038803040515631437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.02228107489645481 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31342294812202454 + ], + "opposite_logprob_tokens": [ + -1.3134229183197021 + ], + "logprob_reward": -0.31342294812202454, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31342294812202454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06818009167909622, + 0.12015275657176971, + 0.054344870150089264, + 0.10336508601903915, + 0.08365223556756973, + 0.029707886278629303, + 0.16008660197257996, + 0.1947050839662552, + 0.11558019369840622, + 0.07022520899772644 + ], + "credit_map": [ + 0.01886383630335331, + 0.005517124198377132, + 0.011560710147023201, + 0.01216774620115757, + 0.004052009899169207, + 0.005503715015947819, + 0.004671557806432247, + 0.005843390244990587, + 0.01506405882537365, + 0.010074322111904621, + 0.014986349269747734, + 0.0149229122325778, + 0.02417515590786934, + 0.015099383890628815, + 0.011230910196900368, + 0.014599661342799664, + 0.007399965077638626, + 0.0068219308741390705, + 0.006485302001237869, + 0.00430147722363472, + 0.008009891025722027, + 0.0066259331069886684, + 0.008115781471133232, + 0.006584589835256338, + 0.01989123411476612, + 0.008777763694524765, + 0.014602859504520893, + 0.011139996349811554, + 0.006491027772426605, + 0.01689506694674492, + 0.007557552773505449, + 0.01800958439707756, + 0.011118113994598389, + 0.00802482757717371, + 0.010291670449078083, + 0.0083059286698699, + 0.008473002351820469, + 0.015675682574510574, + 0.007580850273370743, + 0.01418216060847044, + 0.0047383555211126804, + 0.0022051979321986437, + 0.003155201906338334, + 0.0024673452135175467, + 0.003652791725471616, + 0.005298640578985214, + 0.0028969671111553907, + 0.005293385591357946, + 0.03262699022889137, + 0.02196928858757019, + 0.028090743348002434, + 0.016159912571310997, + 0.015031957998871803, + 0.017852045595645905, + 0.014773298054933548, + 0.013582366518676281, + 0.03412829339504242, + 0.015573606826364994, + 0.04317684471607208, + 0.018793899565935135, + 0.017323238775134087, + 0.03099234588444233, + 0.012891541235148907, + 0.021825317293405533, + 0.016881590709090233, + 0.01167154312133789, + 0.014811396598815918, + 0.013173707760870457, + 0.018097305670380592, + 0.012661020271480083, + 0.00958309881389141, + 0.018700536340475082, + 0.008315484039485455, + 0.0064328438602387905, + 0.007499269675463438, + 0.004725492559373379, + 0.013048815540969372, + 0.008052079938352108, + 0.004836831241846085, + 0.01731438934803009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.31342294812202454 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341667398810387 + ], + "opposite_logprob_tokens": [ + -3.651341676712036 + ], + "logprob_reward": -0.026341667398810387, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341667398810387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09387800842523575, + 0.12002751231193542, + 0.0569966658949852, + 0.2191951423883438, + 0.08550115674734116, + 0.056382156908512115, + 0.08434110134840012, + 0.09539594501256943, + 0.0918084904551506, + 0.09647386521100998 + ], + "credit_map": [ + 0.022732937708497047, + 0.0058944434858858585, + 0.016064876690506935, + 0.02168738655745983, + 0.0057676550932228565, + 0.007938443683087826, + 0.006175451446324587, + 0.007616812363266945, + 0.015269224531948566, + 0.01031036302447319, + 0.013268535025417805, + 0.017203355208039284, + 0.02571920119225979, + 0.014281602576375008, + 0.010585686191916466, + 0.01338954921811819, + 0.0092503996565938, + 0.0059326705522835255, + 0.007200261577963829, + 0.004249634686857462, + 0.010425325483083725, + 0.006723291240632534, + 0.006251939106732607, + 0.006963144987821579, + 0.05203983187675476, + 0.01560248713940382, + 0.029353931546211243, + 0.019326690584421158, + 0.02405868098139763, + 0.030564770102500916, + 0.012354239821434021, + 0.035894520580768585, + 0.018383167684078217, + 0.006328138522803783, + 0.012025713920593262, + 0.004510494880378246, + 0.007588094566017389, + 0.015245201997458935, + 0.006123030092567205, + 0.015297316946089268, + 0.012118597514927387, + 0.004137855023145676, + 0.0073767597787082195, + 0.003895744914188981, + 0.0059769474901258945, + 0.011667280457913876, + 0.004377666395157576, + 0.006831304170191288, + 0.022700663655996323, + 0.01032890286296606, + 0.021799685433506966, + 0.006687100976705551, + 0.005908614490181208, + 0.007250498048961163, + 0.004132932983338833, + 0.005532704759389162, + 0.01709248311817646, + 0.005378627218306065, + 0.02154642529785633, + 0.011379505507647991, + 0.009252269752323627, + 0.014007281512022018, + 0.006006785202771425, + 0.010732561349868774, + 0.009112806059420109, + 0.010426837019622326, + 0.010956904850900173, + 0.00980438757687807, + 0.019219985231757164, + 0.011375188827514648, + 0.0075605809688568115, + 0.01335179153829813, + 0.012365609407424927, + 0.00688821030780673, + 0.011407831683754921, + 0.00898959394544363, + 0.011946395970880985, + 0.011992258951067924, + 0.007620873395353556, + 0.025263087823987007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.026341667398810387 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327805757522583 + ], + "opposite_logprob_tokens": [ + -3.773277997970581 + ], + "logprob_reward": -0.02327805757522583, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327805757522583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10470514744520187, + 0.053527429699897766, + 0.20984117686748505, + 0.12338919937610626, + 0.05621720477938652, + 0.03595894202589989, + 0.2231033891439438, + 0.0582800954580307, + 0.08009643852710724, + 0.05488092079758644 + ], + "credit_map": [ + 0.018352525308728218, + 0.008294359780848026, + 0.01823435351252556, + 0.019752906635403633, + 0.008729090914130211, + 0.011800281703472137, + 0.006201936397701502, + 0.013339696452021599, + 0.007793017663061619, + 0.0047414530999958515, + 0.006952719762921333, + 0.01371701993048191, + 0.006378219462931156, + 0.004132166039198637, + 0.005175539292395115, + 0.0046373032964766026, + 0.02789927087724209, + 0.024350017309188843, + 0.028308134526014328, + 0.016328884288668633, + 0.03728897124528885, + 0.027666423469781876, + 0.021010547876358032, + 0.026988951489329338, + 0.025757713243365288, + 0.01020218338817358, + 0.019769875332713127, + 0.006908041890710592, + 0.01800835132598877, + 0.01809009350836277, + 0.0071541741490364075, + 0.01749878004193306, + 0.01301881019026041, + 0.004897525534033775, + 0.0075094145722687244, + 0.0038641830906271935, + 0.004477997776120901, + 0.007478183601051569, + 0.003969034180045128, + 0.011002060025930405, + 0.007708298042416573, + 0.00241167563945055, + 0.005015498027205467, + 0.002314733574166894, + 0.0040361457504332066, + 0.00815991684794426, + 0.0025238962844014168, + 0.0037887822836637497, + 0.0762835443019867, + 0.007544302847236395, + 0.09810953587293625, + 0.009580572135746479, + 0.012000653892755508, + 0.0068282014690339565, + 0.004056999925523996, + 0.008699608035385609, + 0.007856311276555061, + 0.004579125437885523, + 0.007984872907400131, + 0.00875999964773655, + 0.00821624044328928, + 0.009109167382121086, + 0.0048561315052211285, + 0.006918251980096102, + 0.008621389046311378, + 0.009536362253129482, + 0.010080697014927864, + 0.00817427970468998, + 0.01534844096750021, + 0.00991265568882227, + 0.007002422120422125, + 0.011420197784900665, + 0.008173825219273567, + 0.005088322330266237, + 0.0067614223808050156, + 0.007849449291825294, + 0.003943611867725849, + 0.005983096547424793, + 0.0033356836065649986, + 0.01374551746994257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.02327805757522583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage biplane with a fuselage made of weathered dark green wood and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the biplane shown descending out of control in a spiraling downward motion with one wing lower than the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vibrant circle of faded red, white, and blue visible on the upper wing of the biplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thin plume of oily black smoke billowing from the engine cowling of the aircraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a surreal sky filled with painterly clouds in swirls of sunset orange, deep purple, and golden ochre?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009180859779007733 + ], + "opposite_logprob_tokens": [ + -7.000917911529541 + ], + "logprob_reward": -0.0009180859779007733, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009180859779007733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13895830512046814, + 0.03484893962740898, + 0.05959806963801384, + 0.10026250779628754, + 0.040672238916158676, + 0.05212490260601044, + 0.24279791116714478, + 0.049104008823633194, + 0.0627184808254242, + 0.21891455352306366 + ], + "credit_map": [ + 0.03828737139701843, + 0.009024426341056824, + 0.024936752393841743, + 0.027766816318035126, + 0.009509589523077011, + 0.01183813251554966, + 0.007654412649571896, + 0.009940816089510918, + 0.005983414128422737, + 0.0026259147562086582, + 0.004985545761883259, + 0.009192096069455147, + 0.0032149669714272022, + 0.0031622429378330708, + 0.002509684767574072, + 0.003175076562911272, + 0.017756298184394836, + 0.004777203314006329, + 0.010548676364123821, + 0.004437753930687904, + 0.004483127500861883, + 0.004219299182295799, + 0.0036268061958253384, + 0.009748912416398525, + 0.01812724955379963, + 0.006044795736670494, + 0.015343183651566505, + 0.008895044215023518, + 0.009640870615839958, + 0.015574891120195389, + 0.005682583432644606, + 0.020953895524144173, + 0.011483442969620228, + 0.0025602334644645452, + 0.00588988745585084, + 0.0022383341565728188, + 0.0034849958028644323, + 0.005669285077601671, + 0.002140263095498085, + 0.007205801550298929, + 0.010207862593233585, + 0.003196039004251361, + 0.007160074543207884, + 0.0037362033035606146, + 0.005709765013307333, + 0.012254856526851654, + 0.0035371058620512486, + 0.006323004141449928, + 0.07980180531740189, + 0.00935402698814869, + 0.09671088308095932, + 0.015588797628879547, + 0.011984891258180141, + 0.010638555511832237, + 0.006624761037528515, + 0.012094192206859589, + 0.005620457697659731, + 0.004134297836571932, + 0.005914746783673763, + 0.004916109144687653, + 0.008663835935294628, + 0.00986702460795641, + 0.004058596678078175, + 0.005928943865001202, + 0.011746369302272797, + 0.006657124031335115, + 0.008790891617536545, + 0.006952566560357809, + 0.006295814178884029, + 0.005360090173780918, + 0.005704640410840511, + 0.011210990138351917, + 0.03494687378406525, + 0.01784597709774971, + 0.028987638652324677, + 0.028969045728445053, + 0.011994116008281708, + 0.026046529412269592, + 0.012607092969119549, + 0.05751729756593704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0009180859779007733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_9/index_tensor([7321, 7321])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_9/index_tensor([7321, 7321])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6f14f3e75fcbfe5958efc6426450c097b7c794c3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_9/index_tensor([7321, 7321])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "strikingly cinematic drone shot, captured from a slightly overhead perspective with a smooth, sideways tracking shot, captures a sleek, black horse, its muscles defined and its sturdy frame swiftly galloping across the ancient, winding Great Wall. The first rays of morning sunlight slant down, gilding the horse's silhouette with a dazzling gilded edge. Its long, black mane and tail stream backward in the wind, and its hooves fly, each powerful step striking the broad, weathered, gray bricks. The camera follows the horse's movements, capturing the majestic sweep of the Great Wall as it undulates along the mountain ridgeline, with distant beacon towers looming in the morning mist. The backdrop is a majestic, overlapping mountain range of varying shades of indigo. A light mist lingers in the valleys, and sunlight penetrates the clouds, casting dappled light and shadow across the mountain range. The entire picture is in extremely clear 4K high-definition quality, with warm and layered tones, creating an epic atmosphere that is grand, vigorous, full of freedom and power.", + "seed": 4265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_9/index_7321_seed_4265_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016971701756119728, + "vlm_logprob": -0.016971701756119728, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.016971701756119728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00956051331013441 + ], + "opposite_logprob_tokens": [ + -7.009560585021973 + ], + "logprob_reward": -0.00956051331013441, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.00956051331013441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08589049428701401, + 0.25576794147491455, + 0.2946992516517639, + 0.009359552524983883, + 0.024326875805854797, + 0.04221491888165474, + 0.03573574125766754, + 0.09938280284404755, + 0.087267205119133, + 0.06535524874925613 + ], + "credit_map": [ + 0.014023431576788425, + 0.019247068092226982, + 0.008907640352845192, + 0.007622250355780125, + 0.013510831631720066, + 0.007325653452426195, + 0.007578699383884668, + 0.007674920838326216, + 0.07387667894363403, + 0.02019062452018261, + 0.0859462171792984, + 0.01432285737246275, + 0.012960398569703102, + 0.01678900606930256, + 0.02352013811469078, + 0.00816199742257595, + 0.05861137807369232, + 0.017444469034671783, + 0.1185266450047493, + 0.030268816277384758, + 0.014009363017976284, + 0.025930296629667282, + 0.017559045925736427, + 0.01234923955053091, + 0.0011796345934271812, + 0.0011927745072171092, + 0.0012635656166821718, + 0.0012946539791300893, + 0.00195092405192554, + 0.0009957512374967337, + 0.0007984203984960914, + 0.0006838279077783227, + 0.0027988681104034185, + 0.003019575495272875, + 0.00305054965429008, + 0.002000302542001009, + 0.004008245654404163, + 0.0031572787556797266, + 0.0035995265934616327, + 0.0026925287675112486, + 0.007250753231346607, + 0.002724363235756755, + 0.015614250674843788, + 0.003918317146599293, + 0.0031523345969617367, + 0.0038895101752132177, + 0.0028348767664283514, + 0.002830511424690485, + 0.002440829761326313, + 0.010660148225724697, + 0.0038439896889030933, + 0.0018351974431425333, + 0.01132296398282051, + 0.0022498255129903555, + 0.0015513513935729861, + 0.0018314373446628451, + 0.015561046078801155, + 0.006551654078066349, + 0.020306453108787537, + 0.007635822053998709, + 0.016101734712719917, + 0.008204367011785507, + 0.006178146228194237, + 0.018843578174710274, + 0.010895635932683945, + 0.006944307126104832, + 0.03607466444373131, + 0.007604991551488638, + 0.005803612992167473, + 0.004699563141912222, + 0.0071848370134830475, + 0.008059587329626083, + 0.0032897398341447115, + 0.00273510767146945, + 0.025061212480068207, + 0.0034229138400405645, + 0.018349790945649147, + 0.004119920544326305, + 0.004571668338030577, + 0.0038048955611884594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.00956051331013441 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355484500527382 + ], + "opposite_logprob_tokens": [ + -3.651355504989624 + ], + "logprob_reward": -0.026355484500527382, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355484500527382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04811820760369301, + 0.2952396869659424, + 0.17289577424526215, + 0.04886721819639206, + 0.021804269403219223, + 0.07139519602060318, + 0.028066111728549004, + 0.0514194630086422, + 0.04887530207633972, + 0.21331875026226044 + ], + "credit_map": [ + 0.007636151276528835, + 0.010909562930464745, + 0.00913239549845457, + 0.003062852658331394, + 0.005452117417007685, + 0.00433883722871542, + 0.0026583229191601276, + 0.004927970003336668, + 0.08530189841985703, + 0.02323821932077408, + 0.09984897822141647, + 0.015979133546352386, + 0.015187428332865238, + 0.019021885469555855, + 0.027348458766937256, + 0.009313671849668026, + 0.031155437231063843, + 0.013400029391050339, + 0.06052013114094734, + 0.018687762320041656, + 0.01385753508657217, + 0.01718279719352722, + 0.00979691557586193, + 0.008295181207358837, + 0.0065671843476593494, + 0.004620329011231661, + 0.008975272998213768, + 0.009095732122659683, + 0.005840321071445942, + 0.004046646412461996, + 0.004431798122823238, + 0.005289931781589985, + 0.0029197880066931248, + 0.002811803948134184, + 0.003361009294167161, + 0.001687056734226644, + 0.0035459231585264206, + 0.0033753474708646536, + 0.0022469288669526577, + 0.0018564125057309866, + 0.011639736592769623, + 0.004450393375009298, + 0.029956592246890068, + 0.007176714949309826, + 0.0037719090469181538, + 0.0044457740150392056, + 0.0037548327818512917, + 0.006199243012815714, + 0.002593234647065401, + 0.003943226300179958, + 0.004146726336330175, + 0.0029111644253134727, + 0.005334058310836554, + 0.004042993765324354, + 0.0019507321994751692, + 0.0031439762096852064, + 0.00850659143179655, + 0.003328501945361495, + 0.009012878872454166, + 0.004095786716789007, + 0.009935752488672733, + 0.004439384210854769, + 0.003316775429993868, + 0.008783791214227676, + 0.009240110404789448, + 0.003689757315441966, + 0.01776484213769436, + 0.004074490163475275, + 0.004609441384673119, + 0.003077703295275569, + 0.0042337640188634396, + 0.002185190562158823, + 0.0105351647362113, + 0.009001911617815495, + 0.08079778403043747, + 0.010327775031328201, + 0.06706064194440842, + 0.013608804903924465, + 0.014703773893415928, + 0.007282888516783714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.026355484500527382 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076340362429618835 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.0076340362429618835, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076340362429618835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047522012144327164, + 0.35720279812812805, + 0.12148341536521912, + 0.019879454746842384, + 0.03228479251265526, + 0.03323248773813248, + 0.0163012333214283, + 0.04689646139740944, + 0.09387535601854324, + 0.2313220202922821 + ], + "credit_map": [ + 0.009126920253038406, + 0.013892500661313534, + 0.005111865699291229, + 0.002573483856394887, + 0.008413871750235558, + 0.003770521841943264, + 0.0016664082650095224, + 0.0029664402827620506, + 0.1036287471652031, + 0.02813081629574299, + 0.12052706629037857, + 0.01926957257091999, + 0.01837148144841194, + 0.023196157068014145, + 0.03301037475466728, + 0.011068563908338547, + 0.02322741597890854, + 0.008890781551599503, + 0.04617218300700188, + 0.012545855715870857, + 0.007371034473180771, + 0.010632886551320553, + 0.007153332233428955, + 0.005489925853908062, + 0.0023657719139009714, + 0.0022490210831165314, + 0.0026513540651649237, + 0.0025526368990540504, + 0.0037087365053594112, + 0.0018918850691989064, + 0.0027071295771747828, + 0.0017529191682115197, + 0.0035347784869372845, + 0.003943524789065123, + 0.003750524716451764, + 0.0028059883043169975, + 0.005165289156138897, + 0.004828655160963535, + 0.0045300498604774475, + 0.0037259815726429224, + 0.0045430222526192665, + 0.0033065234310925007, + 0.01095807645469904, + 0.003983574919402599, + 0.00251132738776505, + 0.002880754182115197, + 0.002547097159549594, + 0.0025021121837198734, + 0.001598378294147551, + 0.0030456571839749813, + 0.0020246505737304688, + 0.0016381976893171668, + 0.0025541079230606556, + 0.002599032362923026, + 0.0011691461550071836, + 0.0016720629064366221, + 0.0074495431035757065, + 0.0024877379182726145, + 0.008960412815213203, + 0.0031045216601341963, + 0.006792617030441761, + 0.0038643060252070427, + 0.0030059078708291054, + 0.011231415905058384, + 0.009538319893181324, + 0.005593075882643461, + 0.040058434009552, + 0.00888692494481802, + 0.01127369049936533, + 0.004789323545992374, + 0.006768383085727692, + 0.006967195775359869, + 0.011902524158358574, + 0.01026606559753418, + 0.08389639854431152, + 0.011766805313527584, + 0.0732174888253212, + 0.01726909540593624, + 0.014763926155865192, + 0.008239696733653545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.0076340362429618835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623744197189808 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623744197189808, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623744197189808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02518543414771557, + 0.26779866218566895, + 0.10244741290807724, + 0.02255619317293167, + 0.02129603736102581, + 0.09380839765071869, + 0.06565502285957336, + 0.12407750636339188, + 0.039620570838451385, + 0.23755471408367157 + ], + "credit_map": [ + 0.0034221417736262083, + 0.0036984016187489033, + 0.004089402966201305, + 0.0028515721205621958, + 0.00382319581694901, + 0.002716630697250366, + 0.0021125376224517822, + 0.002471550600603223, + 0.0770203247666359, + 0.020985767245292664, + 0.09079344570636749, + 0.014137287624180317, + 0.013337280601263046, + 0.01756492257118225, + 0.02506968379020691, + 0.008889970369637012, + 0.020489394664764404, + 0.0060822246596217155, + 0.04046987369656563, + 0.010258989408612251, + 0.005756335332989693, + 0.008928938768804073, + 0.006005791015923023, + 0.0044558607041835785, + 0.002700475277379155, + 0.003028260078281164, + 0.0031606024131178856, + 0.002833276754245162, + 0.004536843858659267, + 0.002590036019682884, + 0.0021947582717984915, + 0.0015119402669370174, + 0.002552194520831108, + 0.0025088642723858356, + 0.003625925164669752, + 0.002194174798205495, + 0.002840635133907199, + 0.0033276621252298355, + 0.0021025945898145437, + 0.002143987687304616, + 0.016865748912096024, + 0.0054697017185389996, + 0.038085952401161194, + 0.008848906494677067, + 0.0063367122784256935, + 0.008264580741524696, + 0.004654367454349995, + 0.005282422062009573, + 0.004646274261176586, + 0.018256813287734985, + 0.004606597125530243, + 0.002925615292042494, + 0.023524759337306023, + 0.0058760084211826324, + 0.0027523012831807137, + 0.003066658042371273, + 0.020194316282868385, + 0.007336709648370743, + 0.02209155447781086, + 0.009431786835193634, + 0.021596841514110565, + 0.010571934282779694, + 0.00785914622247219, + 0.024995220825076103, + 0.011772848665714264, + 0.002935401862487197, + 0.007831543684005737, + 0.0024529106449335814, + 0.004093235358595848, + 0.0022696289233863354, + 0.0032101392280310392, + 0.005054864101111889, + 0.01171149779111147, + 0.010129360482096672, + 0.08968190103769302, + 0.011513583362102509, + 0.07482583075761795, + 0.015302777290344238, + 0.016293564811348915, + 0.008096197620034218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.007623744197189808 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033684730529785156 + ], + "opposite_logprob_tokens": [ + -3.408684730529785 + ], + "logprob_reward": -0.033684730529785156, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033684730529785156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09961239993572235, + 0.23287126421928406, + 0.024711530655622482, + 0.01750277727842331, + 0.02404387854039669, + 0.1626114696264267, + 0.01918601803481579, + 0.23489965498447418, + 0.09120996296405792, + 0.09335099905729294 + ], + "credit_map": [ + 0.023636290803551674, + 0.01910233683884144, + 0.011410088278353214, + 0.008477196097373962, + 0.014435048215091228, + 0.008405469357967377, + 0.007375561166554689, + 0.0067704119719564915, + 0.06595989316701889, + 0.0179718267172575, + 0.07865165919065475, + 0.013244896195828915, + 0.011398890987038612, + 0.015551025047898293, + 0.021925117820501328, + 0.008167937397956848, + 0.0031458106823265553, + 0.0031758586410433054, + 0.0031606245320290327, + 0.004481412470340729, + 0.004511595703661442, + 0.0025546939577907324, + 0.0019256570376455784, + 0.0017558764666318893, + 0.0020346897654235363, + 0.002003276254981756, + 0.0028818799182772636, + 0.002544049173593521, + 0.0031521667260676622, + 0.0016386352945119143, + 0.0019290512427687645, + 0.001319027622230351, + 0.0034502940252423286, + 0.0025081424973905087, + 0.003105116542428732, + 0.00176808494143188, + 0.004395041614770889, + 0.0028964418452233076, + 0.0029599913395941257, + 0.002960765967145562, + 0.028159499168395996, + 0.009357401169836521, + 0.06628706306219101, + 0.015783799812197685, + 0.010291595943272114, + 0.013587413355708122, + 0.008586762472987175, + 0.010557929053902626, + 0.0016756854020059109, + 0.003240931546315551, + 0.002892275108024478, + 0.0017772343708202243, + 0.003460475942119956, + 0.0028275579679757357, + 0.001592904096469283, + 0.0017189530190080404, + 0.0394197441637516, + 0.0131703931838274, + 0.04236612096428871, + 0.01650792360305786, + 0.04181511700153351, + 0.019830189645290375, + 0.013575180433690548, + 0.04821498692035675, + 0.00907078292220831, + 0.00643188925459981, + 0.03919261693954468, + 0.008141142316162586, + 0.00633458886295557, + 0.0059100366197526455, + 0.007458316627889872, + 0.008670588955283165, + 0.00485240388661623, + 0.004375871736556292, + 0.030724557116627693, + 0.0045715332962572575, + 0.03030746802687645, + 0.008465331979095936, + 0.005970408208668232, + 0.004083435982465744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.033684730529785156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly cinematic drone shot, captured from a slightly overhead perspective with a smooth, sideways tracking shot, captures a sleek, black horse, its muscles defined and its sturdy frame swiftly galloping across the ancient, winding Great Wall. The first rays of morning sunlight slant down, gilding the horse's silhouette with a dazzling gilded edge. Its long, black mane and tail stream backward in the wind, and its hooves fly, each powerful step striking the broad, weathered, gray bricks. The camera follows the horse's movements, capturing the majestic sweep of the Great Wall as it undulates along the mountain ridgeline, with distant beacon towers looming in the morning mist. The backdrop is a majestic, overlapping mountain range of varying shades of indigo. A light mist lingers in the valleys, and sunlight penetrates the clouds, casting dappled light and shadow across the mountain range. The entire picture is in extremely clear 4K high-definition quality, with warm and layered tones, creating an epic atmosphere that is grand, vigorous, full of freedom and power.", + "seed": 4266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_9/index_7321_seed_4266_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.020704129338264467, + "vlm_logprob": -0.020704129338264467, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.020704129338264467, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01591573655605316 + ], + "opposite_logprob_tokens": [ + -6.390915870666504 + ], + "logprob_reward": -0.01591573655605316, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.01591573655605316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16728106141090393, + 0.09059475362300873, + 0.03309745341539383, + 0.04720588028430939, + 0.06426051259040833, + 0.05679568648338318, + 0.09846194088459015, + 0.28073379397392273, + 0.07742676138877869, + 0.08414214104413986 + ], + "credit_map": [ + 0.030783357098698616, + 0.015594135969877243, + 0.02324371598660946, + 0.011864259839057922, + 0.013248341158032417, + 0.02279728464782238, + 0.024894725531339645, + 0.024855244904756546, + 0.015362229198217392, + 0.008595817722380161, + 0.007192586082965136, + 0.01032757107168436, + 0.01019635982811451, + 0.010944421403110027, + 0.010201938450336456, + 0.017773831263184547, + 0.003934383857995272, + 0.003482415108010173, + 0.003805148880928755, + 0.0051070512272417545, + 0.0028619926888495684, + 0.003872394096106291, + 0.002905943663790822, + 0.007128123659640551, + 0.006897990126162767, + 0.004368280991911888, + 0.005845555104315281, + 0.004538576118648052, + 0.0038728811778128147, + 0.007556957192718983, + 0.005926976446062326, + 0.008198666386306286, + 0.00632475083693862, + 0.006636914797127247, + 0.012509165331721306, + 0.004450493026524782, + 0.004051751922816038, + 0.013563198037445545, + 0.005369475111365318, + 0.011354759335517883, + 0.007089982740581036, + 0.006400660146027803, + 0.011288203299045563, + 0.006005310453474522, + 0.007635236717760563, + 0.0065643624402582645, + 0.006429961882531643, + 0.0053819650784134865, + 0.013190068304538727, + 0.007870752364397049, + 0.010381188243627548, + 0.009553705342113972, + 0.00742900837212801, + 0.0238803680986166, + 0.015841901302337646, + 0.01031495165079832, + 0.043822869658470154, + 0.02440197579562664, + 0.022284189239144325, + 0.026393791660666466, + 0.021482577547430992, + 0.062395576387643814, + 0.030624259263277054, + 0.04932859167456627, + 0.012492965906858444, + 0.007077413145452738, + 0.013884788379073143, + 0.007479980122298002, + 0.005667142570018768, + 0.014084300957620144, + 0.008341447450220585, + 0.008398727513849735, + 0.01110691949725151, + 0.008811745792627335, + 0.006668716669082642, + 0.009296444244682789, + 0.009725943207740784, + 0.009177976287901402, + 0.01735740900039673, + 0.011996988207101822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01591573655605316 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808619990944862 + ], + "opposite_logprob_tokens": [ + -3.529808521270752 + ], + "logprob_reward": -0.029808619990944862, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808619990944862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25894221663475037, + 0.04271746426820755, + 0.046048324555158615, + 0.07234032452106476, + 0.09163932502269745, + 0.07341847568750381, + 0.03664267435669899, + 0.20671556890010834, + 0.04239607974886894, + 0.12913957238197327 + ], + "credit_map": [ + 0.031411152333021164, + 0.028405215591192245, + 0.026696307584643364, + 0.02145436592400074, + 0.027699941769242287, + 0.03697919845581055, + 0.05752431973814964, + 0.028771720826625824, + 0.004042074084281921, + 0.004351303447037935, + 0.00261276145465672, + 0.0037833256646990776, + 0.003861317876726389, + 0.005734981037676334, + 0.006887152325361967, + 0.011444549076259136, + 0.005218018312007189, + 0.0046363151632249355, + 0.00489058718085289, + 0.004851156380027533, + 0.005113258026540279, + 0.005919021554291248, + 0.0057122427970170975, + 0.009707724675536156, + 0.00927711371332407, + 0.008595813997089863, + 0.006860403809696436, + 0.005618199240416288, + 0.0051802764646708965, + 0.015596113167703152, + 0.007501173764467239, + 0.013711233623325825, + 0.007199638057500124, + 0.004918878432363272, + 0.009849727153778076, + 0.004560621455311775, + 0.004370027221739292, + 0.028972042724490166, + 0.005671578925102949, + 0.026096807792782784, + 0.010577335953712463, + 0.009015449322760105, + 0.011419212445616722, + 0.008052804507315159, + 0.009657644666731358, + 0.00694578792899847, + 0.01047280989587307, + 0.00727743050083518, + 0.003877150360494852, + 0.0039663976058363914, + 0.004144818987697363, + 0.0038538023363798857, + 0.0027511424850672483, + 0.006371012888848782, + 0.005656520836055279, + 0.006021827459335327, + 0.03809215873479843, + 0.01490514725446701, + 0.020166201516985893, + 0.019310561940073967, + 0.01253167912364006, + 0.041506387293338776, + 0.02171574905514717, + 0.03848768398165703, + 0.006846637465059757, + 0.004367516376078129, + 0.006433642003685236, + 0.005206020548939705, + 0.004678914323449135, + 0.00464849965646863, + 0.005005537532269955, + 0.005209312774240971, + 0.012790868990123272, + 0.010025110095739365, + 0.01552895363420248, + 0.013787589967250824, + 0.019859295338392258, + 0.025881782174110413, + 0.01674181967973709, + 0.014524155296385288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.029808619990944862 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108264084905386 + ], + "opposite_logprob_tokens": [ + -4.511082649230957 + ], + "logprob_reward": -0.01108264084905386, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108264084905386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15467453002929688, + 0.062090277671813965, + 0.05121859535574913, + 0.048265475779771805, + 0.04912267252802849, + 0.09786549210548401, + 0.1166902557015419, + 0.21393810212612152, + 0.06311473250389099, + 0.143019899725914 + ], + "credit_map": [ + 0.017868103459477425, + 0.017080843448638916, + 0.015253464691340923, + 0.012388955801725388, + 0.014362475834786892, + 0.020800719037652016, + 0.028396308422088623, + 0.028523657470941544, + 0.008020478300750256, + 0.006409363821148872, + 0.0058197141624987125, + 0.007133868057280779, + 0.007521204650402069, + 0.007042415905743837, + 0.007983340881764889, + 0.012159889563918114, + 0.0066481223329901695, + 0.006945361848920584, + 0.006254757754504681, + 0.005332062020897865, + 0.0039006115403026342, + 0.006391949951648712, + 0.004850815050303936, + 0.010894916951656342, + 0.00564468652009964, + 0.004708563443273306, + 0.006963424384593964, + 0.005071430932730436, + 0.005249028094112873, + 0.006107397843152285, + 0.005698846653103828, + 0.00882210023701191, + 0.003501342376694083, + 0.008196961134672165, + 0.00425168639048934, + 0.005114908330142498, + 0.005193531513214111, + 0.009807775728404522, + 0.00476134242489934, + 0.008295123465359211, + 0.011017690412700176, + 0.013071965426206589, + 0.022599102929234505, + 0.010680004023015499, + 0.012950535863637924, + 0.008938525803387165, + 0.009700371883809566, + 0.008907304145395756, + 0.015203794464468956, + 0.004633706994354725, + 0.03492121398448944, + 0.010080174542963505, + 0.004920821636915207, + 0.007493551820516586, + 0.031880125403404236, + 0.007556855212897062, + 0.04827948287129402, + 0.02240077219903469, + 0.01668667234480381, + 0.021275745704770088, + 0.010741762816905975, + 0.022439448162913322, + 0.014348511584103107, + 0.05776570364832878, + 0.009409969672560692, + 0.006514261942356825, + 0.011987121775746346, + 0.00681587029248476, + 0.004790626000612974, + 0.010753629729151726, + 0.006078131031244993, + 0.006765116471797228, + 0.019639482721686363, + 0.011675919406116009, + 0.012267433106899261, + 0.013799522072076797, + 0.013135187327861786, + 0.023649772629141808, + 0.027288811281323433, + 0.021563779562711716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.01108264084905386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634022437036037 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008634022437036037, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634022437036037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13739052414894104, + 0.08347295224666595, + 0.04775619134306908, + 0.0830383375287056, + 0.04614434018731117, + 0.0627281665802002, + 0.08737064898014069, + 0.20608264207839966, + 0.04665485769510269, + 0.19936129450798035 + ], + "credit_map": [ + 0.011694864369928837, + 0.011584985069930553, + 0.011359665542840958, + 0.011111586354672909, + 0.01379923801869154, + 0.0207950621843338, + 0.03346126526594162, + 0.02358386665582657, + 0.009144401177763939, + 0.007519835606217384, + 0.007925605401396751, + 0.011343655176460743, + 0.007446357049047947, + 0.010343643836677074, + 0.009629043750464916, + 0.02012040838599205, + 0.004344375804066658, + 0.003827181411907077, + 0.006045069545507431, + 0.008392044343054295, + 0.004816280212253332, + 0.0065451813861727715, + 0.004651553928852081, + 0.009134508669376373, + 0.008278013207018375, + 0.007056979928165674, + 0.009531158022582531, + 0.008696726523339748, + 0.007604896556586027, + 0.012811237946152687, + 0.015337096527218819, + 0.013722235336899757, + 0.0046515800058841705, + 0.005601388402283192, + 0.006032506935298443, + 0.003627871861681342, + 0.0040048095397651196, + 0.011649919673800468, + 0.003639807226136327, + 0.006936457008123398, + 0.009609303437173367, + 0.006191727705299854, + 0.014886518940329552, + 0.007375854533165693, + 0.005302426405251026, + 0.007113691885024309, + 0.005890845321118832, + 0.006357797421514988, + 0.008630424737930298, + 0.010090576484799385, + 0.009068707004189491, + 0.010245533660054207, + 0.007390755694359541, + 0.013858573511242867, + 0.016545945778489113, + 0.011540127918124199, + 0.051339298486709595, + 0.013581141829490662, + 0.017364969477057457, + 0.01812998577952385, + 0.010410799644887447, + 0.027378709986805916, + 0.016609322279691696, + 0.05126841366291046, + 0.007822695188224316, + 0.004122080281376839, + 0.007563411258161068, + 0.004196607507765293, + 0.0047971876338124275, + 0.007822679355740547, + 0.004522509407252073, + 0.005807689391076565, + 0.0280259121209383, + 0.017043689265847206, + 0.014880594797432423, + 0.01827852614223957, + 0.02063317783176899, + 0.03084971383213997, + 0.03763389214873314, + 0.032015781849622726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.008634022437036037 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038079626858234406 + ], + "opposite_logprob_tokens": [ + -3.2880797386169434 + ], + "logprob_reward": -0.038079626858234406, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038079626858234406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32498931884765625, + 0.025652656331658363, + 0.025549473240971565, + 0.0338982529938221, + 0.041087452322244644, + 0.025503700599074364, + 0.1579679548740387, + 0.21568740904331207, + 0.07957157492637634, + 0.07009221613407135 + ], + "credit_map": [ + 0.01733037456870079, + 0.01290653832256794, + 0.025554388761520386, + 0.024550558999180794, + 0.019885508343577385, + 0.06000390648841858, + 0.10940730571746826, + 0.05535075068473816, + 0.00241238996386528, + 0.0034520693589001894, + 0.002392355352640152, + 0.0028890646062791348, + 0.0022876826114952564, + 0.002918495098128915, + 0.003653974737972021, + 0.005646624602377415, + 0.0022298195399343967, + 0.0030684226658195257, + 0.0025112000294029713, + 0.0030775221530348063, + 0.002050373237580061, + 0.003456635633483529, + 0.002934152027592063, + 0.006221347488462925, + 0.0041155084036290646, + 0.0028156510088592768, + 0.003459793282672763, + 0.0036304891109466553, + 0.004341172520071268, + 0.005334573797881603, + 0.004779765382409096, + 0.005421299487352371, + 0.005188837181776762, + 0.00531117944046855, + 0.005203130189329386, + 0.0035148626193404198, + 0.0029231396038085222, + 0.009260492399334908, + 0.0033542776945978403, + 0.0063315341249108315, + 0.00359225831925869, + 0.005150692071765661, + 0.003634086810052395, + 0.0027269949205219746, + 0.0024227085523307323, + 0.002840970875695348, + 0.0028917579911649227, + 0.0022442303597927094, + 0.017048684880137444, + 0.008571451529860497, + 0.04542097821831703, + 0.01527547650039196, + 0.0076661561615765095, + 0.01144518330693245, + 0.043029457330703735, + 0.00951056182384491, + 0.02540755085647106, + 0.018710482865571976, + 0.0199492946267128, + 0.021708888933062553, + 0.018464230000972748, + 0.06464844197034836, + 0.025916554033756256, + 0.020881948992609978, + 0.012769948691129684, + 0.00629439577460289, + 0.011368770152330399, + 0.0073277512565255165, + 0.005986692849546671, + 0.018162252381443977, + 0.007710444275289774, + 0.009951313957571983, + 0.008707700297236443, + 0.005914319306612015, + 0.006324277725070715, + 0.007860299199819565, + 0.007176126819103956, + 0.014067674055695534, + 0.009468900971114635, + 0.010572919622063637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.038079626858234406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly cinematic drone shot, captured from a slightly overhead perspective with a smooth, sideways tracking shot, captures a sleek, black horse, its muscles defined and its sturdy frame swiftly galloping across the ancient, winding Great Wall. The first rays of morning sunlight slant down, gilding the horse's silhouette with a dazzling gilded edge. Its long, black mane and tail stream backward in the wind, and its hooves fly, each powerful step striking the broad, weathered, gray bricks. The camera follows the horse's movements, capturing the majestic sweep of the Great Wall as it undulates along the mountain ridgeline, with distant beacon towers looming in the morning mist. The backdrop is a majestic, overlapping mountain range of varying shades of indigo. A light mist lingers in the valleys, and sunlight penetrates the clouds, casting dappled light and shadow across the mountain range. The entire picture is in extremely clear 4K high-definition quality, with warm and layered tones, creating an epic atmosphere that is grand, vigorous, full of freedom and power.", + "seed": 4267, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_9/index_7321_seed_4267_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.036332053039222954, + "vlm_logprob": -0.036332053039222954, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.036332053039222954, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03159910440444946 + ], + "opposite_logprob_tokens": [ + -4.531599044799805 + ], + "logprob_reward": -0.03159910440444946, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.03159910440444946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10013037174940109, + 0.23961834609508514, + 0.07322081923484802, + 0.041720617562532425, + 0.07736149430274963, + 0.04917598515748978, + 0.12302489578723907, + 0.04780759662389755, + 0.11074385792016983, + 0.13719597458839417 + ], + "credit_map": [ + 0.010858540423214436, + 0.0069669694639742374, + 0.013434707187116146, + 0.016107885167002678, + 0.013498039916157722, + 0.023968694731593132, + 0.006491558160632849, + 0.008803985081613064, + 0.04104679822921753, + 0.03440167382359505, + 0.01230440940707922, + 0.024888258427381516, + 0.02538415603339672, + 0.07631689310073853, + 0.009868352673947811, + 0.015407804399728775, + 0.014940938912332058, + 0.01729666069149971, + 0.004746459890156984, + 0.005912142805755138, + 0.005198436789214611, + 0.009706654585897923, + 0.005621290300041437, + 0.009798233397305012, + 0.006998884491622448, + 0.007432602345943451, + 0.0027111016679555178, + 0.003818770870566368, + 0.003479048376902938, + 0.010729853063821793, + 0.003356207860633731, + 0.0031941488850861788, + 0.007469585165381432, + 0.009425941854715347, + 0.006317606661468744, + 0.007303665392100811, + 0.01067198347300291, + 0.018706833943724632, + 0.010949951596558094, + 0.006515927612781525, + 0.006819956935942173, + 0.009109027683734894, + 0.007100304588675499, + 0.0035140742547810078, + 0.007734919898211956, + 0.00673710135743022, + 0.003503363113850355, + 0.004657241050153971, + 0.015880951657891273, + 0.017519192770123482, + 0.00802015233784914, + 0.016269640997052193, + 0.011307741515338421, + 0.02437320351600647, + 0.016732599586248398, + 0.012921414338052273, + 0.0060798670165240765, + 0.01222629752010107, + 0.002977808704599738, + 0.00397082231938839, + 0.004718497861176729, + 0.005788146518170834, + 0.0029400731436908245, + 0.009106084704399109, + 0.02988659404218197, + 0.020394394174218178, + 0.010419806465506554, + 0.006664586253464222, + 0.011668577790260315, + 0.011843225918710232, + 0.010588952340185642, + 0.00927772931754589, + 0.023275380954146385, + 0.01445192750543356, + 0.006888031028211117, + 0.009543034248054028, + 0.019084129482507706, + 0.012258325703442097, + 0.014935841783881187, + 0.03675930202007294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.03159910440444946 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904502540826797 + ], + "opposite_logprob_tokens": [ + -2.4640450477600098 + ], + "logprob_reward": -0.08904502540826797, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904502540826797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1573396474123001, + 0.17714199423789978, + 0.09508208185434341, + 0.046519163995981216, + 0.04975035414099693, + 0.053128693252801895, + 0.19494082033634186, + 0.03776249662041664, + 0.10362257063388824, + 0.0847121849656105 + ], + "credit_map": [ + 0.03033401258289814, + 0.012600167654454708, + 0.011640723794698715, + 0.011449940502643585, + 0.0254402756690979, + 0.046518370509147644, + 0.008407778106629848, + 0.010948382318019867, + 0.02792566455900669, + 0.020632274448871613, + 0.010163509286940098, + 0.01735876128077507, + 0.021197596564888954, + 0.051648348569869995, + 0.014331186190247536, + 0.013884651474654675, + 0.014583691954612732, + 0.009068893268704414, + 0.007051586173474789, + 0.012315674684941769, + 0.009666361846029758, + 0.012647159397602081, + 0.013917144387960434, + 0.01583157293498516, + 0.006528121884912252, + 0.008859818801283836, + 0.003981173504143953, + 0.0033803179394453764, + 0.004278198815882206, + 0.013383660465478897, + 0.0029700053855776787, + 0.0031378662679344416, + 0.004451104439795017, + 0.007007829379290342, + 0.004349410999566317, + 0.005306057631969452, + 0.008628287352621555, + 0.007986477576196194, + 0.0075208768248558044, + 0.004500309471040964, + 0.0050360155291855335, + 0.008209330961108208, + 0.006290917284786701, + 0.006134312134236097, + 0.009214342571794987, + 0.0065375398844480515, + 0.007624736987054348, + 0.00408149603754282, + 0.0318603590130806, + 0.030874984338879585, + 0.01590903475880623, + 0.02835042029619217, + 0.014242412522435188, + 0.029577864333987236, + 0.025175798684358597, + 0.01894993707537651, + 0.007856306619942188, + 0.00512459734454751, + 0.002327059395611286, + 0.0028720139525830746, + 0.0036698849871754646, + 0.008747567422688007, + 0.002466877456754446, + 0.0046981871128082275, + 0.010652855038642883, + 0.011154191568493843, + 0.005820993799716234, + 0.007982664741575718, + 0.02904723770916462, + 0.009924747049808502, + 0.0077982572838664055, + 0.021241623908281326, + 0.012690425850450993, + 0.013296032324433327, + 0.009761919267475605, + 0.01180258672684431, + 0.010990923270583153, + 0.008414393290877342, + 0.010258574038743973, + 0.007497322745621204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.08904502540826797 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059713751077652 + ], + "opposite_logprob_tokens": [ + -3.895597219467163 + ], + "logprob_reward": -0.02059713751077652, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059713751077652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1534358710050583, + 0.2573634684085846, + 0.10463855415582657, + 0.10573545098304749, + 0.10145025700330734, + 0.049779292196035385, + 0.08607374876737595, + 0.02078544907271862, + 0.05962909385561943, + 0.061108823865652084 + ], + "credit_map": [ + 0.030630841851234436, + 0.01370411179959774, + 0.010695604607462883, + 0.011159845627844334, + 0.024145357310771942, + 0.04608684405684471, + 0.008772115223109722, + 0.008241157978773117, + 0.04362807050347328, + 0.039664436131715775, + 0.009146040305495262, + 0.019902000203728676, + 0.03193476051092148, + 0.08261846750974655, + 0.01621362939476967, + 0.014256048947572708, + 0.02271680347621441, + 0.01622771844267845, + 0.007628807332366705, + 0.009323280304670334, + 0.011462592519819736, + 0.019689101725816727, + 0.008123541250824928, + 0.00946671050041914, + 0.021538423374295235, + 0.025080107152462006, + 0.006868564058095217, + 0.006132503040134907, + 0.007846038788557053, + 0.024948712438344955, + 0.006305421236902475, + 0.007015680894255638, + 0.011221018619835377, + 0.015429466962814331, + 0.008373714052140713, + 0.007398667745292187, + 0.011500879190862179, + 0.02902212180197239, + 0.00981224700808525, + 0.008692145347595215, + 0.0049288226291537285, + 0.007454439997673035, + 0.00506603391841054, + 0.005091238766908646, + 0.01070912554860115, + 0.006411704234778881, + 0.006110354792326689, + 0.004007576499134302, + 0.010668167844414711, + 0.011623232625424862, + 0.00759915029630065, + 0.013471231795847416, + 0.010435427539050579, + 0.011039611883461475, + 0.012391883879899979, + 0.008845037780702114, + 0.0024055014364421368, + 0.0027921407017856836, + 0.0018147005466744304, + 0.0033920365385711193, + 0.0032106528524309397, + 0.003324997378513217, + 0.0019940868951380253, + 0.0018513319082558155, + 0.006835821084678173, + 0.006301748100668192, + 0.004028096329420805, + 0.004698321223258972, + 0.012065954506397247, + 0.005616907030344009, + 0.010359099134802818, + 0.009723147377371788, + 0.007634973619133234, + 0.009384124539792538, + 0.006873933598399162, + 0.006972618866711855, + 0.010601325891911983, + 0.006504771765321493, + 0.008547945879399776, + 0.0045891315676271915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02059713751077652 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733347196131945 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.006733347196131945, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733347196131945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13334397971630096, + 0.11329694837331772, + 0.1533975750207901, + 0.10024398565292358, + 0.04258930683135986, + 0.12510941922664642, + 0.0711052343249321, + 0.0378122553229332, + 0.14483585953712463, + 0.07826537638902664 + ], + "credit_map": [ + 0.022799460217356682, + 0.01390686072409153, + 0.012504608370363712, + 0.015265698544681072, + 0.019662264734506607, + 0.032773640006780624, + 0.007835617288947105, + 0.008595829829573631, + 0.010583690367639065, + 0.017196834087371826, + 0.006194768939167261, + 0.014945803210139275, + 0.01899215392768383, + 0.021835092455148697, + 0.012841896153986454, + 0.010706707835197449, + 0.03930378332734108, + 0.029351800680160522, + 0.015868958085775375, + 0.011694426648318768, + 0.013649841770529747, + 0.02144455537199974, + 0.01099458709359169, + 0.011089622974395752, + 0.016698194667696953, + 0.024476345628499985, + 0.006062879227101803, + 0.0054522170685231686, + 0.005783189553767443, + 0.030287180095911026, + 0.005822980310767889, + 0.005661003291606903, + 0.003723629517480731, + 0.006707313004881144, + 0.003906699363142252, + 0.005056352820247412, + 0.007606099359691143, + 0.006787524558603764, + 0.005380829330533743, + 0.0034208581782877445, + 0.02060847170650959, + 0.03122810460627079, + 0.012715001590549946, + 0.0088929058983922, + 0.02050146646797657, + 0.016025982797145844, + 0.0081016281619668, + 0.00703585846349597, + 0.012120543979108334, + 0.009206175804138184, + 0.008579079993069172, + 0.009473329409956932, + 0.007534516975283623, + 0.011881568469107151, + 0.005610965192317963, + 0.006699049845337868, + 0.005633615888655186, + 0.008985528722405434, + 0.002212382387369871, + 0.0036772668827325106, + 0.004593139048665762, + 0.004866798408329487, + 0.002267487347126007, + 0.005576035473495722, + 0.01043169666081667, + 0.017689354717731476, + 0.004994579590857029, + 0.009068522602319717, + 0.04840562865138054, + 0.010259175673127174, + 0.009087851271033287, + 0.034899044781923294, + 0.010129119269549847, + 0.011949276551604271, + 0.00777485640719533, + 0.011054087430238724, + 0.013683994300663471, + 0.007802846375852823, + 0.010300756432116032, + 0.005570441018790007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.006733347196131945 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033685650676488876 + ], + "opposite_logprob_tokens": [ + -3.4086856842041016 + ], + "logprob_reward": -0.033685650676488876, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033685650676488876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2324606031179428, + 0.14043556153774261, + 0.14592291414737701, + 0.06062548607587814, + 0.0451287105679512, + 0.12121519446372986, + 0.08997202664613724, + 0.02579985000193119, + 0.028402309864759445, + 0.11003732681274414 + ], + "credit_map": [ + 0.05059410259127617, + 0.014327648095786572, + 0.010915640741586685, + 0.01389935240149498, + 0.04221973195672035, + 0.0771448165178299, + 0.011367767117917538, + 0.011991562321782112, + 0.02402501553297043, + 0.02976146712899208, + 0.005368072539567947, + 0.012926400639116764, + 0.016710933297872543, + 0.032238561660051346, + 0.011819068342447281, + 0.007586054503917694, + 0.024189401417970657, + 0.02896125428378582, + 0.009660929441452026, + 0.015586301684379578, + 0.014292639680206776, + 0.02392113208770752, + 0.012672027572989464, + 0.016639236360788345, + 0.011678149923682213, + 0.011995218694210052, + 0.007198314648121595, + 0.004206227604299784, + 0.0044198655523359776, + 0.012622176669538021, + 0.003370659425854683, + 0.005134876351803541, + 0.004832026083022356, + 0.00602282490581274, + 0.004128334112465382, + 0.004496859386563301, + 0.0058112177066504955, + 0.01025110762566328, + 0.006026395130902529, + 0.003559950739145279, + 0.020971760153770447, + 0.03364541009068489, + 0.011833474971354008, + 0.0072964997962117195, + 0.01756327413022518, + 0.014981144107878208, + 0.007471769116818905, + 0.007451863028109074, + 0.011259977705776691, + 0.02260396257042885, + 0.00597973819822073, + 0.011836882680654526, + 0.007285214029252529, + 0.01638723723590374, + 0.0065158093348145485, + 0.008103216998279095, + 0.003013536799699068, + 0.0026273843832314014, + 0.002347836270928383, + 0.002680970821529627, + 0.0027985936030745506, + 0.006914249621331692, + 0.0016063498333096504, + 0.0038109305314719677, + 0.004850783850997686, + 0.0036979191936552525, + 0.002482013078406453, + 0.0029971918556839228, + 0.004034167155623436, + 0.003194678109139204, + 0.0034872451797127724, + 0.003658313537016511, + 0.019269928336143494, + 0.03481874242424965, + 0.006649438291788101, + 0.007582876365631819, + 0.012959633953869343, + 0.008541211485862732, + 0.006213664542883635, + 0.014001837931573391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.033685650676488876 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly cinematic drone shot, captured from a slightly overhead perspective with a smooth, sideways tracking shot, captures a sleek, black horse, its muscles defined and its sturdy frame swiftly galloping across the ancient, winding Great Wall. The first rays of morning sunlight slant down, gilding the horse's silhouette with a dazzling gilded edge. Its long, black mane and tail stream backward in the wind, and its hooves fly, each powerful step striking the broad, weathered, gray bricks. The camera follows the horse's movements, capturing the majestic sweep of the Great Wall as it undulates along the mountain ridgeline, with distant beacon towers looming in the morning mist. The backdrop is a majestic, overlapping mountain range of varying shades of indigo. A light mist lingers in the valleys, and sunlight penetrates the clouds, casting dappled light and shadow across the mountain range. The entire picture is in extremely clear 4K high-definition quality, with warm and layered tones, creating an epic atmosphere that is grand, vigorous, full of freedom and power.", + "seed": 4268, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000151/group_9/index_7321_seed_4268_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02229436170309782, + "vlm_logprob": -0.02229436170309782, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02229436170309782, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014598343521356583 + ], + "opposite_logprob_tokens": [ + -5.139598369598389 + ], + "logprob_reward": -0.014598343521356583, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.014598343521356583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04962390661239624, + 0.0415206216275692, + 0.044756971299648285, + 0.03897019848227501, + 0.5636146068572998, + 0.14437301456928253, + 0.02835622988641262, + 0.03705023229122162, + 0.018067248165607452, + 0.03366694226861 + ], + "credit_map": [ + 0.005457781255245209, + 0.0052039441652596, + 0.005084196105599403, + 0.009357651695609093, + 0.005607440136373043, + 0.005447057541459799, + 0.006280787754803896, + 0.007185051217675209, + 0.005481034051626921, + 0.00872216746211052, + 0.005594526417553425, + 0.0030070750508457422, + 0.006992965470999479, + 0.006202226504683495, + 0.00249886536039412, + 0.0030217603780329227, + 0.006151181645691395, + 0.007095347158610821, + 0.004244444891810417, + 0.0035769788082689047, + 0.006936559919267893, + 0.006449445616453886, + 0.004979171324521303, + 0.005323847755789757, + 0.00439414894208312, + 0.005052976775914431, + 0.003098319051787257, + 0.0023246307391673326, + 0.006337002385407686, + 0.008709938265383244, + 0.004410848952829838, + 0.004642333369702101, + 0.042507354170084, + 0.20915880799293518, + 0.011863878928124905, + 0.01249485183507204, + 0.2317955642938614, + 0.022210268303751945, + 0.02156480960547924, + 0.01201911922544241, + 0.020180411636829376, + 0.01910749450325966, + 0.031075244769454002, + 0.014635488390922546, + 0.024542782455682755, + 0.01471941638737917, + 0.012490268796682358, + 0.007621909491717815, + 0.0027686816174536943, + 0.004422222264111042, + 0.004264332819730043, + 0.004490268882364035, + 0.003591781947761774, + 0.002483771648257971, + 0.0023836803156882524, + 0.003951491322368383, + 0.004953241907060146, + 0.006031430792063475, + 0.004911986645311117, + 0.003155867103487253, + 0.007653327193111181, + 0.004847127478569746, + 0.00242058839648962, + 0.003076662542298436, + 0.0023674152325838804, + 0.00229329289868474, + 0.00234928447753191, + 0.002127458341419697, + 0.003431173739954829, + 0.0014795964816585183, + 0.0025733697693794966, + 0.0014456574572250247, + 0.005648763384670019, + 0.00668575894087553, + 0.0018706230912357569, + 0.0023939884267747402, + 0.0073087820783257484, + 0.0035128574818372726, + 0.002910592360422015, + 0.0033355799969285727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.014598343521356583 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029811397194862366 + ], + "opposite_logprob_tokens": [ + -3.529811382293701 + ], + "logprob_reward": -0.029811397194862366, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029811397194862366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06774761527776718, + 0.03494884446263313, + 0.05222565680742264, + 0.09001924097537994, + 0.3897457718849182, + 0.16957764327526093, + 0.05729754641652107, + 0.03096853382885456, + 0.06811349838972092, + 0.03935566172003746 + ], + "credit_map": [ + 0.007169770076870918, + 0.008098429068922997, + 0.008074062876403332, + 0.010648046620190144, + 0.006376513745635748, + 0.007312763947993517, + 0.010050376877188683, + 0.010017654858529568, + 0.004113493021577597, + 0.006288762204349041, + 0.0058428761549293995, + 0.0046084257774055, + 0.0027232307475060225, + 0.004291553050279617, + 0.002750893123447895, + 0.0043296110816299915, + 0.009063450619578362, + 0.013023936189711094, + 0.004224292002618313, + 0.003998219966888428, + 0.006125780753791332, + 0.00762450834736228, + 0.0029529996681958437, + 0.005212467163801193, + 0.009945052675902843, + 0.010790400207042694, + 0.0071764150634408, + 0.005508963018655777, + 0.011991865001618862, + 0.020383721217513084, + 0.014093383215367794, + 0.010129433125257492, + 0.029857253655791283, + 0.14464667439460754, + 0.008701515384018421, + 0.009244087152183056, + 0.1603711098432541, + 0.014771715737879276, + 0.013877411372959614, + 0.008275986649096012, + 0.024398427456617355, + 0.022066311910748482, + 0.03854076936841011, + 0.01827416941523552, + 0.03000814840197563, + 0.015709353610873222, + 0.01252971962094307, + 0.008050740696489811, + 0.005482644308358431, + 0.005069192498922348, + 0.012464919127523899, + 0.009719266556203365, + 0.005477237515151501, + 0.003537779441103339, + 0.005428528878837824, + 0.010117977857589722, + 0.005843409337103367, + 0.004488358274102211, + 0.004194493871182203, + 0.0034004475455731153, + 0.005037503316998482, + 0.0028913298156112432, + 0.002522988710552454, + 0.00259000132791698, + 0.00769724790006876, + 0.007963155396282673, + 0.00989437848329544, + 0.01170710101723671, + 0.00957861915230751, + 0.006477021146565676, + 0.01004481315612793, + 0.004751155152916908, + 0.005661957897245884, + 0.005952626466751099, + 0.002890310250222683, + 0.003449021838605404, + 0.009554948657751083, + 0.004624057561159134, + 0.0036951994989067316, + 0.0035275414120405912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.029811397194862366 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647494949400425 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647494949400425, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647494949400425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06443838775157928, + 0.025075098499655724, + 0.08435507118701935, + 0.04761167988181114, + 0.4367901086807251, + 0.17074209451675415, + 0.04290686920285225, + 0.060510918498039246, + 0.025286633521318436, + 0.04228319227695465 + ], + "credit_map": [ + 0.008426925167441368, + 0.00936912838369608, + 0.005913890432566404, + 0.00988784059882164, + 0.005182706285268068, + 0.007364146411418915, + 0.008305069990456104, + 0.009988677687942982, + 0.002812107326462865, + 0.003339549992233515, + 0.0022826152853667736, + 0.0033558837603777647, + 0.0042139641009271145, + 0.004044331144541502, + 0.00249696196988225, + 0.002529681893065572, + 0.009466282092034817, + 0.012964585795998573, + 0.004693626891821623, + 0.0044966391287744045, + 0.02405819296836853, + 0.012453091330826283, + 0.010472874157130718, + 0.005749765783548355, + 0.005776361562311649, + 0.005579345393925905, + 0.005297576077282429, + 0.003842658828943968, + 0.004884958267211914, + 0.009905251674354076, + 0.007521610241383314, + 0.004803912714123726, + 0.03314438462257385, + 0.16614556312561035, + 0.00879664160311222, + 0.009184368886053562, + 0.17586009204387665, + 0.017188088968396187, + 0.016377337276935577, + 0.010093577206134796, + 0.0200540442019701, + 0.02126331813633442, + 0.012739819474518299, + 0.022146400064229965, + 0.01902773790061474, + 0.0288737490773201, + 0.027100039646029472, + 0.019536977633833885, + 0.005402978975325823, + 0.0045669833198189735, + 0.0052792662754654884, + 0.005746016278862953, + 0.006498885806649923, + 0.005609891377389431, + 0.003436516737565398, + 0.006366323679685593, + 0.012186030857264996, + 0.008956542238593102, + 0.007745965383946896, + 0.005332296714186668, + 0.010319112800061703, + 0.008097781799733639, + 0.0039889314211905, + 0.0038842575158923864, + 0.004904449451714754, + 0.0037751994095742702, + 0.002699937205761671, + 0.0017404757672920823, + 0.0060652391985058784, + 0.002250164281576872, + 0.0019206091528758407, + 0.0019305560272186995, + 0.005982160102576017, + 0.0073454598896205425, + 0.0038194216322153807, + 0.004837821237742901, + 0.008402629755437374, + 0.004148532170802355, + 0.0038855262100696564, + 0.00386163592338562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.008647494949400425 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780345484614372 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780345484614372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780345484614372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06832751631736755, + 0.055830780416727066, + 0.058816783130168915, + 0.06077920272946358, + 0.30840280652046204, + 0.2881622612476349, + 0.04476224631071091, + 0.04051312059164047, + 0.03331835940480232, + 0.04108688607811928 + ], + "credit_map": [ + 0.0063746110536158085, + 0.008865101262927055, + 0.006960978265851736, + 0.013996896333992481, + 0.006514932494610548, + 0.008379204198718071, + 0.007777276914566755, + 0.0094585120677948, + 0.005183004308491945, + 0.007173512130975723, + 0.008848401717841625, + 0.007936875335872173, + 0.007425653748214245, + 0.006146895233541727, + 0.006800699047744274, + 0.00631573935970664, + 0.008877107873558998, + 0.011583229526877403, + 0.005454917438328266, + 0.004508179146796465, + 0.007932189851999283, + 0.009263938292860985, + 0.004415209870785475, + 0.0067820074036717415, + 0.010373348370194435, + 0.007284281775355339, + 0.00810606312006712, + 0.008134493604302406, + 0.0063166627660393715, + 0.007634381297975779, + 0.007394533604383469, + 0.00553543958812952, + 0.02390831708908081, + 0.11167410016059875, + 0.0072127147577703, + 0.00767501862719655, + 0.12681780755519867, + 0.012299206107854843, + 0.011813969351351261, + 0.0070016770623624325, + 0.038583334535360336, + 0.038280244916677475, + 0.06204737722873688, + 0.0299410130828619, + 0.046780698001384735, + 0.02863987348973751, + 0.027277015149593353, + 0.016612716019153595, + 0.004735772963613272, + 0.005340768955647945, + 0.007266528904438019, + 0.005352372769266367, + 0.0076602003537118435, + 0.0061419373378157616, + 0.002941422164440155, + 0.005323245655745268, + 0.006250569596886635, + 0.006662039086222649, + 0.005137935746461153, + 0.004706938285380602, + 0.007747584953904152, + 0.004206439945846796, + 0.0031403081957250834, + 0.0026613047812134027, + 0.004739467054605484, + 0.00463212002068758, + 0.0051775178872048855, + 0.003897052723914385, + 0.0052900174632668495, + 0.003088804893195629, + 0.003255132120102644, + 0.0032382484059780836, + 0.006391545757651329, + 0.006353404838591814, + 0.0034033043775707483, + 0.00660026166588068, + 0.0039820196107029915, + 0.004749579355120659, + 0.004191520158201456, + 0.005415250547230244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.009780345484614372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek black horse with a defined muscular frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the horse galloping across the ancient, winding Great Wall made of broad, weathered gray bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the horse have a long black mane and tail streaming backward in the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene captured from a slightly overhead perspective with a smooth sideways tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the backdrop a mountain range of varying shades of indigo with distant beacon towers in the morning mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048634227365255356 + ], + "opposite_logprob_tokens": [ + -3.0486342906951904 + ], + "logprob_reward": -0.048634227365255356, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048634227365255356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04228883609175682, + 0.06029040366411209, + 0.08023115247488022, + 0.14444883167743683, + 0.2392406165599823, + 0.09652700275182724, + 0.10751170665025711, + 0.030036484822630882, + 0.12894344329833984, + 0.0704815536737442 + ], + "credit_map": [ + 0.005257897544652224, + 0.00673733139410615, + 0.005043082404881716, + 0.005973154678940773, + 0.0049396418035030365, + 0.005070772487670183, + 0.004167312290519476, + 0.005099642090499401, + 0.009056349284946918, + 0.014806622639298439, + 0.006682056002318859, + 0.005347054451704025, + 0.00683135399594903, + 0.009198581799864769, + 0.0039286185055971146, + 0.004439766984432936, + 0.012535268440842628, + 0.016454607248306274, + 0.008613549172878265, + 0.009039299562573433, + 0.00896709505468607, + 0.010958370752632618, + 0.0037249126471579075, + 0.009938042610883713, + 0.02870945818722248, + 0.017174813896417618, + 0.02074390836060047, + 0.020045574754476547, + 0.015272554941475391, + 0.014879022724926472, + 0.013716349378228188, + 0.013907147571444511, + 0.018192153424024582, + 0.0901431068778038, + 0.0054180496372282505, + 0.00638533802703023, + 0.09544087946414948, + 0.009038659743964672, + 0.009462780319154263, + 0.0051596383564174175, + 0.01228317990899086, + 0.011406853795051575, + 0.010976685211062431, + 0.010670007206499577, + 0.009959829039871693, + 0.013370092958211899, + 0.013519256375730038, + 0.014341100119054317, + 0.013014916330575943, + 0.015544149093329906, + 0.011906886473298073, + 0.010739174671471119, + 0.01842641457915306, + 0.010163088329136372, + 0.008514180779457092, + 0.019202889874577522, + 0.006811077706515789, + 0.005112263839691877, + 0.004039826337248087, + 0.002530488884076476, + 0.004516774322837591, + 0.0034976936876773834, + 0.001664685900323093, + 0.001863674959167838, + 0.02695554681122303, + 0.02100849337875843, + 0.010158635675907135, + 0.006060084328055382, + 0.03716423362493515, + 0.011224516667425632, + 0.00863894633948803, + 0.007732984609901905, + 0.009297747164964676, + 0.010430148802697659, + 0.005331473425030708, + 0.007967307232320309, + 0.01485363021492958, + 0.010101192630827427, + 0.00552662368863821, + 0.006973435636609793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.048634227365255356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file